Access to this full-text is provided by Wiley.
Content available from CAAI Transactions on Intelligence Technology
This content is subject to copyright. Terms and conditions apply.
Received: 4 April 2023
-
Revised: 2 November 2024
-
Accepted: 26 December 2024
-
CAAI Transactions on Intelligence Technology
DOI: 10.1049/cit2.70000
ORIGINAL RESEARCH
Enhancing patient rehabilitation predictions with a hybrid
anomaly detection model: Density‐based clustering and
interquartile range methods
Murad Ali Khan
1
|Jong‐Hyun Jang
2
|Naeem Iqbal
3,4
|Harun Jamil
5
|
Syed Shehryar Ali Naqvi
5
|Salabat Khan
3,6
|Jae‐Chul Kim
2
|Do‐Hyeun Kim
1,7
1
Department of Computer Engineering, Jeju
National University, Jeju, Republic of Korea
2
Electronics and Telecommunications Research
Institute, Daejeon, Republic of Korea
3
Big Data Center, Jeju National University, Jeju,
Republic of Korea
4
School of Electronics, Electrical Engineering and
Computer Science, Queen’s University Belfast,
Belfast, UK
5
Department of Electronics Engineering, Jeju
National University, Jeju, Republic of Korea
6
Department of Computer Science, COMSATS
University Islamabad, Attock, Pakistan
7
Advanced Technology Research Institute, Jeju
National University, Jeju, Republic of Korea
Correspondence
Do‐Hyeun Kim.
Email: kimdh@jejunu.ac.kr
Funding information
Jeju National University
Abstract
In recent years, there has been a concerted effort to improve anomaly detection tech-
niques, particularly in the context of high‐dimensional, distributed clinical data. Analysing
patient data within clinical settings reveals a pronounced focus on rening diagnostic
accuracy, personalising treatment plans, and optimising resource allocation to enhance
clinical outcomes. Nonetheless, this domain faces unique challenges, such as irregular data
collection, inconsistent data quality, and patient‐specic structural variations. This paper
proposed a novel hybrid approach that integrates heuristic and stochastic methods for
anomaly detection in patient clinical data to address these challenges. The strategy
combines HPO‐based optimal Density‐Based Spatial Clustering of Applications with
Noise for clustering patient exercise data, facilitating efcient anomaly identication.
Subsequently, a stochastic method based on the Interquartile Range lters unreliable data
points, ensuring that medical tools and professionals receive only the most pertinent and
accurate information. The primary objective of this study is to equip healthcare pro-
fessionals and researchers with a robust tool for managing extensive, high‐dimensional
clinical datasets, enabling effective isolation and removal of aberrant data points.
Furthermore, a sophisticated regression model has been developed using Automated
Machine Learning (AutoML) to assess the impact of the ensemble abnormal pattern
detection approach. Various statistical error estimation techniques validate the efcacy of
the hybrid approach alongside AutoML. Experimental results show that implementing
this innovative hybrid model on patient rehabilitation data leads to a notable enhance-
ment in AutoML performance, with an average improvement of 0.041 in the R2score,
surpassing the effectiveness of traditional regression models.
KEYWORDS
articial intelligence, articial neural network, computational intelligence, data analysis, data mining, data
privacy, data protection
1
|
INTRODUCTION
Recently, the amount of data has surged across various aspects
of daily life, including healthcare [1], nance [2], smart
manufacturing [3] etc. The growth of data in healthcare has
increased from 500 petabytes in 2012 to 25,000 petabytes in
2022 [4]. However, a massive amount of healthcare data con-
tains abnormalities, ambiguities, and inconsistencies because of
high dimensionality and distributed environments. Therefore,
it remains a pivotal and comprehensive research area within
This is an open access article under the terms of the Creative Commons Attribution License, which permits use, distribution and reproduction in any medium, provided the original work is
properly cited.
© 2025 The Author(s). CAAI Transactions on Intelligence Technology published by John Wiley & Sons Ltd on behalf of The Institution of Engineering and Technology and Chongqing
University of Technology.
CAAI Trans. Intell. Technol. 2025;1–24. wileyonlinelibrary.com/journal/cit2
-
1
data mining (DM) to tackle strange data behaviour and help
medical persons by giving meaningful information from the
data [5, 6]. Researchers need to be able to detect anomalous
data to attain the necessary knowledge to aid in making better
decisions about data analysis. The difculties posed by large,
multidimensional datasets are considerable, as such data almost
always contains some form of abnormal information, regard-
less of its size [7]. Identifying anomalies poses a signicant
challenge in DM because of the limited availability of labelled
data for atypical instances. Consequently, many methods used
for anomaly detection are inherently unsupervised [8]. An
example of such a task is identifying instances that substantially
deviate from the rest of a dataset. However, creating a uni-
versally applicable mathematical deviation formulation remains
an elusive objective across datasets and scenarios.
The unsupervised nature of anomaly detection often leads
to a notable disparity between statistically unusual and genu-
inely meaningful instances, especially in healthcare settings.
The challenge is compounded by the increased complexity and
volume of data driven by technological advancements in
medical equipment, computer hardware, and other domains.
Traditional methods such as ensembling, sub‐sampling [9, 10],
deep learning [11], and density peak clustering [12] are
continuously being adapted to handle the large‐scale distribu-
tion and high dimensionality of such data. However, detecting
meaningful abnormalities remains intricate, necessitating the
development of more accurate, generalisable, and efcient
anomaly detection frameworks.
Furthermore, the relevance of anomalies can vary with
context, inuenced by factors, such as seasonal trends and
specic time units [13]. For instance, a heart rate that appears
abnormal post‐exercise might be expected considering the
activity, similar to uctuations in the sales of seasonal products
or electricity usage patterns. Current anomaly detection
methods must adapt to these variations, employing specialised
mechanisms to differentiate between true anomalies and con-
textually explained deviations [14, 15]. The ongoing research
strives to rene these techniques, ensuring that anomalies
identied in clinical data do not lead to removing vital infor-
mation, which could otherwise result in signicant healthcare
implications [16, 17]. This highlights the critical need for
continued innovation in anomaly detection methodologies that
are both sensitive to the nuances of clinical data and robust
against false positives.
Explainable abnormal detection has gained traction, utilising
sophisticated models like copulas to enhance interpretability. For
instance, the COPOD framework, introduced in one study [18],
innovatively estimates tail probabilities using empirical copulas,
improving the precision of anomaly detection in clinical con-
texts. Moreover, the identication of multivariate abnormalities
has been explored through various methodologies, such as
clinical trial‐based mechanisms [19] and robust statistical tech-
niques. The Hawkins dataset, for example, serves as a funda-
mental resource in several studies, like the one by Gao et al. [20],
which compares the Max‐Eigen Difference technique against
traditional Robust and Mahalanobis distances for detecting
multivariate anomalies. Further contributions include the
unmasking of multivariate anomalies using novel approaches as
discussed by Southworth [21], and the application of non‐
parametric multivariate depth functions to reliably distinguish
between abnormal and normal data [22]. These methodologies
underscore the ongoing efforts to rene anomaly detection
processes, making them more applicable and explainable within
clinical settings.
Existing methods have several limitations, as most are
specically designed to address anomalous data points using
supervised data. These methods, however, necessitate a large
volume of training observations, which are costly to gather and
analyse. As a result, the performance of these methods is
insufciently effective in handling anomalous data instances in
mission‐critical applications, such as the medical eld. More-
over, the curse of dimensionality limits these methods' ability
to nd patterns in low‐dimensional data. The effectiveness of
these methods in demonstrating the need and dependability of
handling anomalous data patterns in the healthcare industry
has not been conrmed. Therefore, a strong and trustworthy
abnormal patterns detection approach is needed to handle
patient clinical data and identify anomalous data patterns by
giving medical professionals just pertinent patient information.
To prociently identify abnormal data observations within
extensive sets of patient clinical data, this study presents a
hybrid method for detecting abnormal data, which combines
heuristic and stochastic techniques. Furthermore, notable
contributions of the proposed hybrid scheme are listed below:
�Create a sophisticated data preparation module that can
perform various tasks to enhance the accuracy and calibre of
patient data obtained.
�Design a hybrid abnormal data detection methodology that
combines stochastic and heuristic techniques to identify
strange data behaviour in patient data, enabling medical
professionals to assist patients by supplying pertinent in-
formation exclusively.
�The effectiveness of the proposed hybrid abnormal data
detection technique in identifying strange data behaviour in
patient data is assessed utilising the Automated Machine
Learning (AutoML) paradigm.
�A range of statistical techniques are incorporated to check
the effectiveness and emphasise the importance of the
suggested study investigation.
2
|
RELATED WORK
Anomaly detection, dened as identifying deviations from
expected behaviour, plays a crucial role in machine learning
(ML) and medical data analysis, particularly in improving the
accuracy and reliability of predictive models. This section or-
ganises the related work into three main categories: Anomaly
Detection in ML, Anomaly Detection in Medical and Clinical
Applications, and Statistical and Computational Techniques for
Anomaly Detection.
2
-
KHAN ET AL.
2.1
|
Anomaly detection in machine learning
In the ML eld, anomalies are typically considered rare oc-
currences or events that deviate from standard patterns,
prompting closer inspection due to their unusual nature. Ex-
amples include sudden changes in activity, text errors, or
abrupt temperature shifts [23–25]. Within this context, ML and
DM approaches have evolved to address anomaly detection
using methods, such as Bayesian Networks, Naive Bayes, and
Decision Trees, which are valued for their simplicity and
effectiveness [26]. For example, the K‐Nearest Neighbour al-
gorithm has gained popularity for its robust results in
proximity‐based learning, although it requires considerable
computational resources and storage [27].
Further advancements in ML‐based anomaly detection
involve distinguishing genuine issues from false positives.
Zheng et al. [28] developed a dataset to train models specif-
ically for identifying vulnerabilities, helping to lter out false
alarms in static analysis and allowing developers to prioritise
actual security issues efciently. Anomaly detection has also
extended to using Generative Adversarial Networks (GANs)
and Autoencoders in medical applications, such as optical
coherence tomography for anomaly detection [29] and early‐
stage kidney transplant rejection [30]. However, signicant
challenges remain particularly the high false‐positive rates and
the need for consistent denitions and criteria for what con-
stitutes an anomaly [31–33]. These drawbacks underscore the
importance of rening ML techniques to enhance practical
applicability and reduce computational complexity.
2.2
|
Anomaly detection in medical and
clinical applications
Anomaly detection plays an essential role in medical and
clinical data analysis, where early detection of unusual patterns
can improve patient outcomes and support timely in-
terventions. Minimum spanning tree‐based methods, intro-
duced by Li et al. [34, 35], handle anomalies in complex
medical datasets, enhancing diagnostic accuracy. In other
research, Gijzel et al. [36] provided insights into predicting
resilience and physical strength through various tests, though
further investigation is needed to address limitations in clinical
implementation.
Meta‐analyses, such as the one by Ji et al. [37], highlight the
reduced relative risk of cardiac events through data‐driven
methods. However, ndings may not generalise across all
populations. Gevaert et al. [38] proposed personalised exercise
interventions, considering individual differences such as he-
reditary factors. In contrast, Morris et al. [39] reported that
exercise‐based rehabilitation improved the quality of life for
patients with pulmonary hypertension. Similarly, Babaei et al.
[40] explored the use of body‐worn accelerometers for reha-
bilitation, Dibben et al. [41] reviewed cardiac rehabilitation for
coronary heart disease, and Cheng et al. [42] and Gu et al. [43]
conducted meta‐analyses on pulmonary rehabilitation (PR) for
patients with interstitial lung disease.
Numerous studies support the utility of PR for improving
exercise capacity and alleviating breathlessness in chronic res-
piratory conditions, including [44–47]. In another example,
Rasheed et al. [48] examined ML methods for early seizure
prediction through electroencephalography signals, while Li
et al. [49] explored various ML techniques for classifying car-
diovascular diseases. Other applications of anomaly detection
in medical contexts include tracking patient postures with long
short‐term memory (LSTM) models [50] and using fusion‐
based approaches to handle data sparsity, as in the study by
Yang et al. [51]. Zhang et al. [52] demonstrated convolutional
neural network (CNN) effectiveness in binary classication of
imbalanced datasets. Liu et al. [53] and Zhao et al. [54] lever-
aged ML models for rehabilitation and patient data analysis,
respectively.
2.3
|
Statistical and computational
techniques for anomaly detection
Statistical methods are foundational in anomaly detection due
to their efciency and simplicity. Thresholding is a widely used
approach where predened limits indicate anomalies, as
demonstrated by Amores et al. [55], who used an intelligent
trash bin with gas sensors to monitor food quality. Another
standard method, Analysis of Variance, determines signicant
differences between groups and is valuable for identifying
unusual patterns [56].
More complex frameworks, such as the Expectation‐
Maximisation ensemble unsupervised clustering driven by su-
pervised learning, are tailored to clustering tasks and have
demonstrated superior accuracy in psychiatric disease analysis
by approximately þ1.9 points over other methods [57]. Li et al.
[58] proposed the Poincare distance metric for anomaly
detection in electronic medical records, enhancing reliability. In
contrast, Amanullah et al. [59] highlighted the prognostic sig-
nicance of extra‐aortic valvular cardiac abnormalities in aortic
stenosis patients, linking severity with mortality and adverse
events. Peng et al. [60] utilised a latent correlation method for
mood and anxiety disorders analysis, which, despite challenges
in establishing a shared latent structure, underscored the need
for targeted approaches in psychopathology.
Finally, statistical techniques play a critical role in rehabil-
itation data analysis. Models like CNN and LSTM help monitor
patient recovery [50, 52], while risk assessment models provide
valuable tools for clinical prognosis [56]. These studies high-
light the versatility of statistical and computational techniques
in detecting anomalies, though limitations such as scalability
and adaptability to high‐dimensional data persist.
3
|
PROPOSED ABNORMALITY
PREDICTION FRAMEWORK
This section outlines the suggested approach for this article.
Detecting abnormalities in patient's rehabilitation data neces-
sitates a sensitive and reliable framework for identifying
KHAN ET AL.
-
3
anomalous data with enhanced predictive performance. The
healthcare domain currently grapples with the intricate chal-
lenges of sophisticated technologies, making detecting
abnormal data particularly difcult. Given the potential risks to
human life, decisions within this domain are crucial.
3.1
|
Overview model
Unusual data patterns signify abnormal data stream behaviour
diverging from the actual data stream. Employing heuristic and
stochastic methods becomes imperative for identifying and
managing these anomalies. This study uses the Hyperparameter
Optimisation (HPO‐based) optimised Density‐Based Spatial
Clustering of Applications with Noise (DBSCAN) as a
heuristic approach to discern and segregate abnormal data
patterns, bolstering data reliability. Additionally, stochastic
techniques encompass statistical methods like mean, thresh-
olding, range, and level shift to pinpoint abnormal data pat-
terns. This study adopts an augmented approach to enhance
precision by integrating heuristic and stochastic techniques for
more effective abnormality detection. The illustrated model
addressing abnormal data is depicted in Figure 1.
3.2
|
Proposed hybrid strategy architectural
overview
One of the biggest problems in dealing with abnormalities is
the lack of data resources in any eld, especially in the medical
FIGURE 1 Recommended hybrid model for anomaly detection. AutoML, automated machine learning; DBSCAN, density‐based spatial clustering of
applications with noise; LGB, light GB; MAE, mean absolute error; MAPE, mean absolute percentage error; MSE, mean squared error; RF, random forest;
RMSE, root mean squared error.
4
-
KHAN ET AL.
domain. Unfortunately, to evaluate the proposed model, we
have a small number of records in the original dataset with
highly unreliable and missing values. So, rst, the available data
is passed through the preprocessing step to cope with unreli-
ability in data and make it in a form that any prediction model
can further use.
An adaptive imputation method lls in missing values to
enhance data quality and consistency. This study utilises K‐
nearest neighbors and zero‐based imputation techniques for
continuous attributes, while predictive analytics is employed to
ll missing cells in categorical attributes. After the pre-
processing stage, the data, although prepared, is insufcient for
practical training of deep learning models. Consequently, there
is a need to enhance this limited dataset into a more substantial
synthetic dataset to ensure precise model training. This study
applies the GAN augmentation technique to expand the
dataset signicantly, generating synthetic samples from actual
data to train the model accurately.
The primary objective of this study is to predict abnormal
records in clinical data for effective patient rehabilitation. To
assess the inuence of eliminating abnormal data, the initial
experiment is conducted on the original augmented data, and
the performance of each model is recorded. Subsequently, each
dataset undergoes DBSCAN based on the parameters outlined
in this study [61], evaluating the predictive models' perfor-
mance. HPO is utilised to optimise DBSCAN, selecting the
best parameters for efcient abnormal data detection in each
dataset. An advanced hybrid algorithm for abnormal pattern
detection, integrating Heuristic and Stochastic algorithms, is
proposed to eliminate strange data patterns from the actual
augmented data. This hybrid model incorporates optimised
DBSCAN with the Interquartile Range (IQR) to detect and
eradicate anomalous records across all four datasets. The same
models assess data reliability after removing aberrant records
from the clinical data. The overall prediction performances are
then compared as depicted in Figure 2, to assess the inuence
of removing strange patterns from the actual data with the help
of the suggested model.
3.3
|
Detailed ow model
The IQR is a statistical measure used to identify outliers in a
dataset. To calculate the IQR, the data is rst sorted and then
divided into quartiles: the lower quartile (Q1), median (Q2),
and upper quartile (Q3). The IQR is the range between Q3 and
Q1, which represents the middle 50% of the data. The lower
and upper bounds for detecting outliers are then dened as
Q1 −1.5IQR and Q3 þ1.5IQR, respectively. Any data points
outside these bounds are considered abnormal or outliers as
illustrated in Figure 3. This approach is benecial for skewed
distributions or datasets with non‐normal distributions, as it
focuses on the spread of the middle 50% of the data, effec-
tively ltering out extreme values that could distort statistical
analyses.
3.4
|
The hybrid strategy proposed for
abnormality detection
Techniques for detecting abnormalities focus on identifying
deviations from standard behavioural patterns. The aim is to
establish a baseline for what constitutes normal behaviour and
dene a corresponding normality region. To achieve this, an
algorithm is devised for segregating abnormal data from
standard data, as depicted in the Algorithm 1. The suggested
FIGURE 2 Proposed design schematic for identifying anomalies in clinical data. DBSCAN, density‐based spatial clustering of applications with noise; IQ,
interquartile; MAE, mean absolute error; MAPE, mean absolute percentage error; MSE, mean squared error; RMSE, root mean squared error.
KHAN ET AL.
-
5
hybrid approach for identifying abnormal data points involves
several steps.
At the outset, the data collected is subjected to a cleansing
process, which involves removing redundant and unchanging
attributes to reduce computational load. Following this, a
hybrid method is applied that integrates predictive analytics
with the k‐NN algorithm to address the issue of incomplete
feature values. The prepared data then undergoes analysis
through the DBSCAN algorithm to identify initial anomalies in
patients' clinical and exercise data. The initial DBSCAN
implementation employs parameters outlined in Ref. [61], and
the results are reported based on these predened settings.
Furthermore, the HPO method is applied to ne‐tune
parameters specic to each dataset, ensuring optimal func-
tionality of the DBSCAN algorithm. After ne‐tuning, the
DBSCAN algorithm is combined with a stochastic approach
based on the IQR to pinpoint data points that deviate from the
norm within each dataset. This IQR‐focused stochastic tech-
nique functions through a series of steps: initially, the rst
inter‐quantile range of the data is computed. Subsequently, this
value is multiplied by a constant factor of 1.5. The result is then
added to the third quantile, marking any data point exceeding
this sum as abnormal. Conversely, the same multiplied value is
subtracted from the rst quantile, and any data point falling
below this difference is labelled as abnormal. These steps
collectively enable the detection of potential anomalies within
the dataset.
This methodology effectively isolates and eliminates atyp-
ical (outlier) data from the processed dataset, thereby offering
crucial insights to healthcare professionals for evaluating pa-
tient recovery subsequently.
Algorithm 2presents a detailed protocol for evaluating the
efcacy of the newly formulated strategy for detecting ab-
normalities, with a specic focus on leveraging AutoML. The
algorithm unfolds in several critical stages:
The process begins with enhancing the data quality
through preprocessing, as outlined in Algorithm 1. Post‐
preprocessing, the dataset is bifurcated into training (denoted
as Strain) and testing (denoted as Stest ) subsets. Next, AutoML is
utilised to improve the learning process, especially in a
specialised regression task. AutoML's growing prominence is
attributed to its ability to automate the development of ML
models, making it an indispensable tool. The subsequent phase
involves selecting the most effective regression model based on
AutoML's performance in the training stage. The chosen
model is subsequently utilised to evaluate the effectiveness of
the hybrid approach in predicting health‐related metrics.
AutoML dramatically improves the process by automating
the selection and optimisation of ML models. This efcient
method enhances workow productivity and saves a signicant
FIGURE 3 Working of interquartile range abnormal detection.
Algorithm 1 Hybrid approach for identifying
atypical data points within clinical and exercise-
related patient data.
6
-
KHAN ET AL.
amount of time and resources that manual model optimisation
would typically require. Transitioning to the fth phase of the
process, the performance of the model selected by AutoML is
evaluated using the Stest samples. Key metrics, including Mean
Absolute Error (MAE), Mean Squared Error (MSE), Root
Mean Squared Error (RMSE), Coefcient of Determination
(R2 score), Root Mean Squared Logarithmic Error (RMSLE),
and Mean Absolute Percentage Error (MAPE) are utilised to
gauge the accuracy of predictions, offering valuable insights
into the model's efcacy during testing. In the nal stage, a
comparison is made between various baseline models,
including XGBoost (XGB), Light GB (LGB), and Random
Forest (RF), and the optimal regression model derived from
AutoML. This comparison is essential for evaluating the rela-
tive success of the hybrid approach in detecting anomalies
within the selected clinical data.
4
|
DATA PREPARATION AND
ANALYSIS
4.1
|
Data collection
To assess the effectiveness of the developed hybrid method, we
have sourced genuine patient data from Jeju National Uni-
versity (JNU) in South Korea. The research concentrates on
the clinical records of patients receiving treatment for Adhesive
Articular Cystitis. The dataset procured includes vital infor-
mation, such as patient demographics, disease conditions, risk
assessments, and functional capacity evaluations. The
demographic section covers basic patient information,
including ID, name, gender etc. The disease status is detailed
with attributes, such as current health condition, physical ac-
tivity level, specic exercise routines, disease identication,
primary symptoms, date of symptom onset, appointment dates
etc. The segment on risk prediction encompasses a variety of
data points, including both personal and medical histories. For
functional evaluation, the dataset includes measurements like
grip strength (right and left), arm curl tests (right and left), and
the Push‐up test. In this dataset, specic elds are not labelled
with conventional names but are instead assigned a distinct
‘data code’ for each entry. Utilising this authentic clinical data
from actual patients is crucial for validating the effectiveness of
the proposed hybrid model in detecting anomalies.
4.2
|
Data preprocessing
Data preparation focuses on cleaning and transforming data
into a reliable format to enhance the discovery of hidden
patterns and insights within the data. This preprocessing phase
involves multiple steps aimed at achieving high‐quality,
dependable data. These steps include detecting and eliminating
duplicate rows within the dataset. Furthermore, non‐essential
and static features are discarded to mitigate computational
and storage burdens. The processing of multi‐valued attributes
ensures each attribute maintains atomicity. Hybrid data impu-
tation techniques, integrating k‐NN and predictive methods,
are applied to ll in missing attribute values. Data trans-
formation processes are utilised to adapt attribute values into a
Algorithm 2 A rened regression methodology using AutoML is employed to assess the efcacy of the
suggested hybrid model in detecting anomalies.
KHAN
ET AL.
-
7
format that machines can interpret. Additionally, data nor-
malisation methods are considered to standardise values within
a specic range to ensure uniformity in attribute values.
4.3
|
Data analysis
Figure 4displays the outcomes of a Pearson correlation anal-
ysis performed on the functional evaluation attributes. This
technique calculates the linear relationship between any two
specied attributes using the Pearson correlation coefcient,
which varies from −1 to 1. A coefcient of 1 denotes a perfect
positive linear correlation, while −1 indicates a perfect negative
linear correlation, and 0 implies no linear correlation exists
between the variables. Pearson's ris widely utilised in statistics
to measure the degree of linear correlation between variables.
The matrix shown in Figure 4is a tabular arrangement
displaying the correlation coefcients among different variable
sets. It is important to highlight that the diagonal elements of
the matrix are consistently assigned a value of 1, indicating that
the correlation of any variable with itself is inherently perfect.
Using correlation analysis, we pinpoint features in the
dataset that exhibit positive and negative correlations. In the
experimental setup, we focus solely on the features with pos-
itive correlations and deliberately omit those with negative
correlations from the dataset. In the specic dataset, only three
features—FE0015, FE0014, and FE0007—demonstrate
negative correlations with the target attributes, as depicted in
Figure 4. This strategic selection focuses on features that
favourably correlate with the target attributes for further
analysis.
4.4
|
Data augmentation
ML models, particularly deep learning models, demand a
robust training process reliant on extensively annotated data-
sets. However, acquiring, storing, and processing such large
datasets can be prohibitively expensive. Furthermore, the ac-
curacy of these models hinges on access to a substantial
number of high‐quality labelled training data records, often
numbering in the tens of thousands or more. Consequently,
there is a pressing need for a solution to augment small
datasets into larger synthetic datasets, enabling the develop-
ment of generalised models.
This study introduces an innovative method named the
GAN‐assisted Time GAN augmentation technique. This
approach expands small datasets into larger ones by producing
synthetic samples derived from actual data. Additionally,
Figure 5visually represents the impact of data augmentation
using GAN, illustrating the increase in instances before and after
augmentation. Initially, the number of instances for each dataset
is around one thousand, and after augmentation, the number of
records for each dataset is increased to ten thousand. This
augmentation process signicantly enhances the dataset size,
contributing to more effective and generalised model training.
5
|
OUTCOMES OF THE STUDY
5.1
|
Testing setup
Table 1provides an overview of the diverse tools and tech-
nologies utilised to craft the suggested data and predictive
analytics framework. Python emerges as the principal
FIGURE 4 Identifying key features through the application of
correlation analysis.
FIGURE 5 Actual and synthetic data for the experiments.
TABLE 1Tools and technologies used in experimentation.
Component of the
system Details
Operating system Windows 11 operating system
Main memory 96 GB RAM
CPU 12th Generation IntelR CoreTM i9‐12900K at
3.20 GHz
Programming environment Python 3
Development IDE PyCharm expert edition
Database MS Excel, MySQL
Core toolkits Pandas, Sklearn, Seaborn, Matplotlib,
ydatasynthetic, ctgan, AutoML, DBSCAN etc.
Abbreviation: IDE, integrated development environment.
8
-
KHAN ET AL.
programming language for conducting experiments in this
research. Several essential Python libraries, such as Sklearn,
ydata_synthetic, ctgan, Pycaret, and ADTK, have been used.
Furthermore, Microsoft Excel has been utilised to store both
raw and processed data.
5.2
|
Results
This section details the outcomes of utilising DBSCAN for
anomaly detection across four separate datasets: the Borg
rating of perceived exertion (RPE), the TUG, the Romberg
Test, and the Berg Balance Scale (BBS). We follow a systematic
approach that includes initial visualisation, the same parameters
used by Braei and Wagner [61], parameter optimisation
through grid search, and a hybrid approach incorporating the
DBSCAN with optimised parameters and the IQR for
enhanced abnormality detection. A basic operational illustra-
tion of DBSCAN is presented in Algorithm 3, showcasing the
fundamental procedures of the DBSCAN algorithm. Algo-
rithm 3takes three key parameters:
�Data: The dataset under examination.
�ε(Epsilon): The radius within which the algorithm searches
for neighbouring data points around each data point.
�Minimum Points (MinPts): This refers to the requisite count
of data points that must exist in the ε‐vicinity of a given data
point to qualify as a core point.
The algorithm proceeds as follows: It iterates through each
unvisited data point in the dataset. Each data point checks if
the number of data points within the ε‐neighbourhood of that
point exceeds the MinPts threshold. If the condition is met, a
new cluster is created, and the expandCluster function is called
to grow the cluster by recursively adding connected data
points. The data point is labelled as noise if the condition is
not met.
In the DBSCAN algorithm, the ε‐neighbourhood NεðPÞof
a point Pis dened as the set of all points qwhose distance
from Pis less than or equal to εas shown in Equation (1):
NεðPÞ ¼ fq2Data∣distanceðP;qÞ≤εg ð1Þ
A point Pis considered a corepoint if the number of points
within its ε‐neighbourhood is greater than or equal to the
minimum number of points MinPts as shown in Equation (2):
jNεðPÞj ≥MinPts ð2Þ
5.2.1
|
Simple data visualisation
We commence by gaining insights into the datasets through
simple data visualisation techniques as shown in Figure 6. The
visualisation helps understand the data's inherent structure and
provides a basis for selecting appropriate parameters for the
DBSCAN algorithm.
5.2.2
|
DBSCAN with reference paper
parameters
Next, we apply DBSCAN to the datasets using parameters
from the referenced paper [61]. The DBSCAN algorithm aims
to identify dense regions in the data, treating data points with
insufcient density as noise. It assigns labels to data points
based on density and connectivity, using parameters dened in
the referenced paper [61]. A visual representation of abnormal
detection in four datasets with DBSCAN with reference paper
parameters is shown in Figure 7.
5.2.3
|
DBSCAN parameter optimisation with
grid search
To enhance the effectiveness of DBSCAN, we perform
parameter optimisation using grid search. We explore a range
of εvalues and MinPts values to nd the optimal combination
that best suits the characteristics of each dataset. The grid
FIGURE 6 A simple visualisation of four data before abnormal
detection. BBS, Berg Balance Scale.
Algorithm 3 DBSCAN algorithm.
KHAN
ET AL.
-
9
search is a systematic method for hyperparameter tuning, and it
aims to maximise the silhouette score, which quanties the
separation between clusters and the density of the data points
within each cluster. The grid search optimisation is used to nd
the optimal parameters εand MinPts by maximising the
silhouette score Ss. The optimisation process is formulated as
Equation (3):
ðε∗
;MinPts∗Þ ¼ arg max
ε;MinPts Ssðε;MinPtsÞ ð3Þ
where ε∗and MinPts∗denote the optimal values of εand
MinPts, respectively, and Ssðε;MinPtsÞis the silhouette score
corresponding to the chosen values of εand MinPts. The grid
search iterates over predened ranges of εand MinPts values,
systematically evaluating the silhouette score to identify the
best parameters. Mathematically, the silhouette score (S) is
dened in Equation (4):
Ss ¼y−x
maxðx;yÞð4Þ
where xrepresents the mean distance between a specic data
point and other points in its cluster, conversely, yindicates the
mean distance from the same data point to the nearest points
in the closest neighbouring cluster. The resulting abnormal
points using optimised DBSCAN parameters are shown in
Figure 8for four datasets.
5.2.4
|
Hybrid abnormal detection with optimised
DBSCAN and IQR
In addition to optimising DBSCAN parameters, a hybrid
approach is introduced that leverages the IQR for further
abnormality detection. After clustering data points using
DBSCAN with optimised parameters, we identify outliers by
considering points outside the IQR boundaries. The IQR is
calculated as follows Equation (5):
IQR ¼Q3−Q1ð5Þ
In this context, Q1 denotes the lower quartile of the
dataset, whereas Q3 indicates the upper quartile. Outliers are
those data points outside the interval ½Q1−k∗IQR;Q3þ
k∗IQR�, with kbeing a constant the user determines.
The hybrid abnormal detection method combines the
optimised DBSCAN‐based clustering and the IQR‐based ab-
normality detection. The hybrid abnormality score His
computed by combining the silhouette score Sfrom the
DBSCAN clustering and the abnormality score from IQR,
weighted by W. The nal equation for hybrid abnormality
detection is given in Equation (6):
H¼α⋅Sþ ð1−αÞ⋅IQRscore ð6Þ
In this equation, Sis the silhouette score obtained from the
optimised DBSCAN algorithm, and IQRscore is the outlier
score derived from the IQR method. The parameter α2 ½0;1�
is a user‐dened weighting factor that controls the relative
importance of DBSCAN clustering ðSÞand IQR‐based outlier
detection ðIQRscoreÞ. A higher αemphasises the DBSCAN
clustering, while a lower αgives more weight to the IQR‐based
detection. A visual illustration of abnormal detection for all
four datasets using Hybrid Abnormal Detection with Opti-
mised DBSCAN and IQR is given in Figure 9.
6
|
VALIDATION RESULTS AND
ANALYSIS
This segment assesses how well the regression models forecast
the dependent variable within the research. We gauge the
models' accuracy by juxtaposing their predictions against the
FIGURE 7 Visualisation of abnormal detection of density‐based
spatial clustering of applications with noise with reference paper
parameters. BBS, Berg Balance Scale.
FIGURE 8 Visualisation of abnormal detection with optimised
density‐based spatial clustering of applications with noise parameters.
10
-
KHAN ET AL.
real‐world results. To assess the models' effectiveness, we use
evaluation metrics, including the R2 score and other error
metrics, to determine the quality of each experiment.
6.1
|
Validation metrics
In ML, regression is used to nd relationships between
dependent and independent variables. Regression, in its simple
denition, is a process of predicting discrete values like prices,
consumption, ratings etc.
In clinical practice, medical experts rely on various per-
formance indicators to assess the reliability and accuracy of
predictive models. Metrics such as MAE, MSE, RMSE, R2,
RMSLE, and MAPE provide valuable insights into the model's
predictive capabilities. For instance, MAE, MSE, and RMSE
quantify the magnitude of prediction errors, whereas R2in-
dicates the proportion of variance explained by the model.
RMSLE is particularly useful for datasets with a wide range of
values, and MAPE measures prediction accuracy relative to the
actual values. By understanding and interpreting these in-
dicators, medical experts can make informed judgements about
the model's performance and suitability for clinical decision
support systems and patient care management.
�R2 Score: The performance of a regression‐based ML
model is often assessed using the R2 score. This serves as an
indicator of the model's tting accuracy. This metric, also
known as the coefcient of determination, is represented as
Rsquared. It provides insight into the variance explained by
a dataset through the model's predictions. The R2 score is
determined using the following Equation (7):
R2¼1−SSR
SSt ð7Þ
In the given equation, SSR is the squared differences be-
tween the observed and predicted values, while SSt denotes
the total squared deviation of each data point from the
mean. The R2 score ranges from 0 to 1 and is typically
presented as a percentage.
�MAE: Mean Absolute Error is a metric that calculates the
average of the absolute differences between the observed
values and the predictions generated by the model. It's a
straightforward measure of how close the model's pre-
dictions are to the actual data. Taking an example where
input and output data are available for the model will help
in better understanding and linear regression will be used
to determine the best‐t line. To calculate model MAE,
we must determine if the model made a false outcome,
also known as an error. After nding the difference be-
tween actual data and predicted values, we calculate the
MAE for the overall dataset. MAE is determined by
summing the errors across all instances and then dividing
this total by the number of cases. It signies the loss
value, and the goal is to reduce it as much as possible.
The MAE value is computed using Equation (8) as shown
below:
MAE ¼1
xX
x
i¼1
jyi−^
yij ð8Þ
In this equation, yirepresents the actual data, ^
yirepresents
the model's predictions, and xis the total number of in-
stances. Minimising MAE is a common goal in many
regression and predictive modelling tasks.
�MSE: Mean Squared Error is frequently employed when
assessing model performance, differing subtly from MAE.
The primary variation in MSE lies in its computation of
errors, squaring the difference between predicted and actual
values, thereby accentuating the size of the errors. MAE and
MSE differ primarily because MSE measures the mean
squared errors, whereas MAE measures the absolute errors.
MSE has the advantage of taking squared values, which
prevents the cancellation of terms due to negative values.
The computation of MSE is expressed in the following
Equation (9):
MSE ¼1
nX
n
i¼1xi−yi�2ð9Þ
Within Equation (9), yidenotes the values produced by the
trained model, xicorresponds to the ground truth values in
the dataset, and nshows the number of instances in the
data.
�RMSE: is especially valuable for assessing the dispersion of
errors relative to the actual values. A reduced RMSE in-
dicates that the predictions made by the model are closely
aligned with the actual data points. Conversely, a higher
RMSE indicates a more signicant average error. The pro-
cedure for calculating the RMSE value is detailed in
Equation (10).
FIGURE 9 Visualisation of hybrid abnormal detection with optimised
density‐based spatial clustering of applications with noise and interquartile
range.
KHAN ET AL.
-
11
RMSE ¼
1
mX
m
j¼1aj−bj�2
v
u
u
tð10Þ
In Equation (10), bjrepresents the value generated by the
model, ajdenotes the actual value found in the data, and m
indicates the total number of instances within the dataset.
�MAPE: MAPE is a metric used to assess the accuracy of a
forecasting or predictive model. It quanties the average
percentage difference between the predicted and actual
values. MAPE is expressed as a percentage, making it easy to
understand and interpret. The formula for MAPE is given in
Equation (11).
MAPE ¼1
mX
m
i¼1
�
�
�
�
Ai−Pi
Ai�
�
�
�
�100 ð11Þ
where mis the total number of data points. Airepresents
the actual value for data point i.Pirepresents the predicted
(forecasted) value for data point i.
6.2
|
Validation results
In this section, we are going to visualise the actual and pre-
dicted values of four diseases data and check the inuence of
abnormal detection using DBSCAN with referenced paper
parameters [61], optimised DBSCAN parameters, and the last
one, which is optimised DBSCAN incorporated with IQR
which is the proposed abnormal detection of this study.
6.2.1
|
Original data results
In the rst case, the original data results are illustrated to check
the data pattern without removing any abnormal instances
from the data. In this scenario, four disease data, Borg RPE,
TUG, Romgberg test, and BBS, are predicted using LGB,
XGB, RF, and AutoML. Figure 10, have four different part
Figure 10a–d. Figure 10a shows the actual and predicted values
for Borg RPE data using LGB, XGB, RF, and AutoML.
Ground truth data is visualised with a solid blue line, LGB
prediction with the dash‐dotted orange line, XGB prediction
with a dotted green line, RF prediction with a dash‐dotted red
line, and AutoML prediction with a dotted purple line. In the
case of Borg RPE, as shown in Figure 10a, all four models
have good prediction, but some blue solid line is still visible on
some peek instances. Furthermore, in Figure 10b, prediction
for TUG disease data shows moderate results because many
predictions go beyond the solid line. In the case of Romgberg
test disease, Figure 10c models have a good prediction for the
middle values but a limitation for peek values prediction where
a solid line is visible. For the BBS disease data, Figure 10d
models missed the blue line and predicted beyond blue in some
predictions.
6.2.2
|
Prediction after DBSCAN with paper
parameters
In this section, we explore the impact of DBSCAN with
parameters outlined in the study [61] to detect and remove
abnormal instances from the dataset. The aim is to assess
FIGURE 10 Actual and predicted data visualisation before abnormal detection. AutoML, automated machine learning; BBS, Berg Balance Scale; RF,
random forest.
12
-
KHAN ET AL.
how this preprocessing step inuences the performance of
four prediction models on four different types of data: Borg
RPE, TUG, Romberg Test, and BBS. Figure 11 showcases the
actual and predicted data for each of the four diseases after
applying DBSCAN with the specied parameters from the
study [61].
It is worth noting that, as indicated in the analysis, the
results are not as promising as before. The DBSCAN algo-
rithm with the chosen parameters has removed a signicant
amount of non‐abnormal data, leading to a decline in predic-
tion performance. This decline can be observed in the actual
and predicted graphs as shown in Figure 11a–d. This section
highlights the impact of the DBSCAN preprocessing step,
emphasising the trade‐off between removing abnormal in-
stances and potentially sacricing predictive accuracy.
6.2.3
|
Prediction after optimised DBSCAN
This section investigates the effect of utilising optimised
DBSCAN parameters obtained through grid search optimisa-
tion to detect and remove abnormal instances from the dataset.
The objective is to assess the impact of this improved pre-
processing step on the predictive performance of four ML
models for four distinct types of disease: Borg RPE, TUG,
Romberg Test, and BBS. Figure 12a–d display the actual and
predicted data for each of the four diseases after applying
DBSCAN with optimised parameters. This optimised
DBSCAN preprocessing step improves the prediction
performance compared to the previous experiments. However,
it is worth noting that in some instances, the models still fall
short of accurately matching the actual data (blue line).
This section underscores the positive impact of optimising
DBSCAN parameters in improving data preprocessing.
Nevertheless, the models may still face challenges in capturing
certain nuances in the data, as observed in the actual and
predicted graphs specially in Figure 12a.
6.2.4
|
Prediction after proposed hybrid
abnormal detection
Figure 13 illustrates the actual and predicted data for four
different diseases following the application of the proposed
hybrid abnormal detection technique.
These visual representations vividly illustrate the efciency
of the proposed hybrid anomaly detection method. The actual
and predicted data closely overlap, underscoring the superior
performance of the detection approach. This precise congru-
ence between the actual and predicted values showcases the
hybrid model's capability to enhance data quality by adeptly
identifying and eliminating abnormal data points.
Furthermore, this research reveals that AutoML, the sec-
ondary proposed model, demonstrates enhanced predictive
capabilities compared to leading models, such as LGB, XGB,
and RF. The remarkable alignment between actual and pre-
dicted values in the graphs underscores AutoML's capacity to
capture intricate patterns and relationships within the data.
FIGURE 11 Actual and predicted data visualisation after AD with density‐based spatial clustering of applications with noise using referenced paper
parameters [61]. AutoML, automated machine learning; BBS, Berg Balance Scale; RF, random forest.
KHAN ET AL.
-
13
In conclusion, the adoption of the hybrid abnormal
detection technique, which combines optimised DBSCAN
parameters with IQR, results in a substantial improvement in
predictive accuracy. Furthermore, the integration of AutoML
as one of the predictive models underscores its potential as a
leading model selection and HPO tool.
FIGURE 12 Actual and predicted data visualisation after AD with optimised density‐based spatial clustering of applications with noise. AutoML,
automated machine learning; BBS, Berg Balance Scale; RF, random forest.
FIGURE 13 Actual and predicted data visualisation after AD with proposed Hybrid AD model. AutoML, automated machine learning; BBS, Berg Balance
Scale; RF, random forest.
14
-
KHAN ET AL.
6.3
|
Validation analysis
6.3.1
|
Borg RPE: VP0004 analysis
In the comprehensive analysis of the Borg RPE dataset
(VP0004), various machine‐learning models were assessed for
their performance in different abnormal detection mechanisms
as presented in Table 2. AutoML consistently stood out as the
top performer, maintaining the lowest MAE, MSE, and RMSE
values and the highest R2 (0.9713) in the original VP0004
dataset. Additionally, LGB delivered competitive results,
particularly in RMSE (0.0474) and R2 (0.9599). These trends
persisted when applying DBSCAN with paper‐referenced pa-
rameters to VP0004, with AutoML achieving an MAE of
0.031, RMSE of 0.051, and an R2 of 0.858, while LGB and
XGB models maintained competitive positions. Similarly, in
scenarios involving VP0004 with DBSCAN and optimised
parameters, AutoML consistently outperformed its peers with
an MAE of 0.015, RMSE of 0.033, and a robust R2 of 0.974.
The superiority of AutoML persisted in the context of VP0004
with optimised DBSCAN parameters combined with IQR,
where it achieved the lowest MAE (0.013), RMSE (0.030), and
an impressive R2 (0.998). Notably, LGB and XGB models
maintained their competitive performance throughout these
experiments.
In summary, while the ML models signicantly validated
the abnormal detection mechanisms, the primary focus re-
mains on evaluating them. The optimal choice may vary,
contingent on dataset characteristics and specic requirements.
However, the hybrid approach, combining optimised
DBSCAN parameters with IQR, consistently emerges as the
most effective mechanism for anomaly detection within the
VP0004 dataset.
6.3.2
|
TUG: VP0005 analysis
In the extensive analysis of the VP0005 TUG dataset, various
ML models were evaluated to determine their effectiveness in
different abnormal detection mechanisms as detailed in
Table 3. Across the board, AutoML consistently emerged as
the top performer, consistently recording the lowest MAE,
MSE, RMSE and the highest R2 value (0.948) on the original
VP0005 data. Additionally, the LGB model showcased
competitive results, particularly excelling in RMSE (0.052) and
R2 (0.950). These trends remained consistent when applying
DBSCAN with paper‐referenced parameters to VP0005.
AutoML achieved an MAE of 0.0598, RMSE of 0.0598, and an
R2 of 0.948, whereas LGB and XGB models maintained
competitive positions.
Similarly, in scenarios involving VP0005 with DBSCAN
and optimised parameters, AutoML consistently outperformed
its peers, recording an MAE of 0.019, RMSE of 0.040, and a
robust R2 of 0.958. The LGB model also delivered competitive
results, especially regarding RMSE (0.047) and R2 (0.941).
Furthermore, VP0005 with optimised DBSCAN parameters
TABLE 2Performance comparison of ML models in abnormal detection for VP0004 Borg RPE dataset.
AD techniques Models MAE MSE RMSE R2 RMSLE MAPE
VP0004 LGB 0.0222 0.0026 0.0474 0.9599 0.0306 0.0778
XGB 0.0234 0.0019 0.0422 0.9696 0.0273 0.0742
RF 0.0257 0.0024 0.0469 0.9633 0.0306 0.0759
AutoML 0.0216 0.0019 0.0415 0.9713 0.0267 0.0657
VP0004 (DBSCAN paper parameters) LGB 0.035 0.005 0.071 0.805 0.045 0.10
XGB 0.034 0.004 0.058 0.820 0.040 0.093
RF 0.037 0.005 0.071 0.798 0.046 0.105
AutoML 0.031 0.003 0.051 0.858 0.038 0.085
VP0004 (DBSCAN optimise parameters) LGB 0.018 0.0012 0.040 0.965 0.023 0.070
XGB 0.017 0.0011 0.035 0.970 0.021 0.067
RF 0.019 0.0013 0.041 0.963 0.024 0.072
AutoML 0.015 0.001 0.033 0.974 0.020 0.065
VP0004 hybrid (optimal DBSCAN þIQR) LGB 0.015 0.0012 0.035 0.976 0.018 0.057
XGB 0.014 0.0010 0.032 0.982 0.016 0.055
RF 0.016 0.0013 0.036 0.975 0.018 0.060
AutoML 0.013 0.0009 0.030 0.998 0.015 0.052
Note: Bold values in the results Tables showing the Proposed model highest results.
Abbreviations: AutoML, automated machine learning; BBS, Berg Balance Scale; DBSCAN, density‐based spatial clustering of applications with noise; IQR, interquartile range; LGB, light
GB; MAE, mean absolute error; MAPE, mean absolute percentage error; MSE, mean squared error; RF, random forest; RMSE, root mean squared error; RMSLE, root mean squared
logarithmic error; RPE, rating of perceived exertion; XGB, XGBoost.
KHAN ET AL.
-
15
and the integration of the IQR proved advantageous for
AutoML, which stood out as the top performer in this sce-
nario, achieving the lowest MAE (0.019), RMSE (0.034), and
an impressive R2 (0.993). Remarkably, the LGB model also
maintained competitive performance in this context.
The primary objective throughout this analysis is to select
the best abnormal detection mechanism for the VP0005 TUG
dataset, considering factors such as dataset characteristics and
specic requirements. While ML models played a signicant
role in validating these mechanisms, the focus remains pri-
marily on evaluating them. The choice may vary depending on
the specic dataset. Still, the hybrid approach that combines
optimised DBSCAN parameters with IQR consistently
emerges as the most effective mechanism for anomaly detec-
tion within the VP0005 dataset.
6.3.3
|
Romgberg test: VP0006 analysis
Table 4presents a comprehensive evaluation of detecting ab-
normalities within the Romberg Test dataset (VP0006). The
models consistently demonstrate varying degrees of effective-
ness across different scenarios, including variations in
DBSCAN parameters and the integration of the IQR. AutoML
consistently achieves the lowest RMSE and highest R2 values,
indicating its robustness in capturing the underlying patterns
within the VP0006 dataset. Furthermore, the results
underscore the importance of parameter optimisation, as
models generally exhibit improved performance when utilising
optimised DBSCAN parameters.
In the scenario involving optimised DBSCAN parameters,
signicant performance enhancements are observed across all
models. For example, LGB achieves an RMSE of 0.034 and an
R2 of 0.962, whereas AutoML demonstrates even better per-
formance with an RMSE of 0.032 and an impressive R2 of
0.974. Moreover, integrating the IQR with optimised
DBSCAN parameters further improves abnormal detection
accuracy. In this hybrid approach, AutoML stands out with an
RMSE of 0.032 and an exceptional R2 of 0.999, indicating a
near‐perfect t to the data. These ndings underscore the
effectiveness of advanced techniques and parameter optimi-
sation in enhancing abnormal detection mechanisms for the
VP0006 dataset.
6.3.4
|
BBS: VP0007 analysis
In the case of abnormalities detection within the BBS: VP0007
dataset, different congurations of DBSCAN parameters, and
the integration of the IQR, the models exhibit distinct per-
formance characteristics. When DBSCAN parameters undergo
optimisation, substantial ameliorations in model efcacy are
discernible across the board. For instance, LGB achieves an
impressive RMSE of 0.043 and an R2 of 0.952, whereas
TABLE 3Performance comparison of ML models in abnormal detection for VP0005 TUG dataset.
Data Models MAE MSE RMSE R2 RMSLE MAPE
VP0005 LGB 0.026 0.003 0.052 0.950 0.045 0.0521
XGB 0.027 0.003 0.051 0.952 0.044 0.068
RF 0.028 0.003 0.054 0.948 0.048 0.0598
AutoML 0.0598 0.0598 0.0598 0.948 0.044 0.0519
VP0005 (DBSCAN paper parameters) LGB 0.028 0.004 0.065 0.925 0.036 0.090
XGB 0.027 0.003 0.052 0.936 0.033 0.085
RF 0.030 0.004 0.068 0.920 0.038 0.095
AutoML 0.023 0.023 0.046 0.954 0.030 0.075
VP0005 (DBSCAN optimise parameters) LGB 0.022 0.0022 0.047 0.941 0.030 0.075
XGB 0.021 0.0019 0.044 0.951 0.028 0.070
RF 0.024 0.0024 0.049 0.936 0.032 0.080
AutoML 0.019 0.0016 0.040 0.958 0.026 0.065
VP0005 hybrid (optimal DBSCAN þIQR) LGB 0.021 0.0016 0.040 0.963 0.017 0.055
XGB 0.020 0.0013 0.036 0.975 0.016 0.052
RF 0.022 0.0018 0.042 0.961 0.018 0.057
AutoML 0.019 0.0012 0.034 0.993 0.015 0.050
Note: Bold values in the results Tables showing the Proposed model highest results.
Abbreviations: AutoML, automated machine learning; BBS, Berg Balance Scale; DBSCAN, density‐based spatial clustering of applications with noise; IQR, interquartile range; LGB, light
GB; MAE, mean absolute error; MAPE, mean absolute percentage error; MSE, mean squared error; RF, random forest; RMSE, root mean squared error; RMSLE, root mean squared
logarithmic error; XGB, XGBoost.
16
-
KHAN ET AL.
AutoML exhibits even more commendable performance with
an RMSE of 0.037 and an exceptional R2 of 0.968. Further-
more, the amalgamation of the IQR with optimised DBSCAN
parameters yields further enhancements in abnormal detection
accuracy. In this hybrid methodology, AutoML stands out with
an RMSE of 0.030 and an outstanding R2 of 0.997, signifying a
near‐perfect t to the data. These ndings underscore the ef-
cacy of advanced techniques and parameter optimisation in
augmenting abnormal detection mechanisms for the VP0007
dataset. The detailed results of BBS: VP0007 data are shown in
Table 5.
7
|
DISCUSSION
This section thoroughly examines the impact and efcacy of
the proposed research, emphasising its contributions and im-
plications. We highlight the signicance of the ndings
compared to prior abnormal data detection research, focusing
on the adverse effects of abnormal data on model perfor-
mance. Removing abnormal data points enhances model
robustness by reducing noise and bias, leading to improved
generalisation and more accurate predictions. Empirical evi-
dence supports the observed performance improvements,
emphasising the importance of data preprocessing steps such
as outlier detection. We also assess performance metrics like
the R2 score and MAPE to evaluate model effectiveness.
7.1
|
Performance analysis in terms of R2
score
The initial facet of the investigation revolves around scruti-
nising the discernible performance improvements quantied
through the R2 score. Figure 14 serves as a visual aid to
articulate the trajectory of enhancement in predictive accuracy,
offering a comprehensive view of the advancement achieved
by the proposed methodology. Figure 14a shows the
comparative analysis for Borg RPE data; in the initial evalua-
tion using various models on the original data, we established a
baseline with R2 scores for LGB (0.959), XGB (0.969), RF
(0.963), and AutoML (0.971). The subsequent introduction of
DBSCAN with parameters from Study [2] for abnormal data
removal resulted in a discernible impact on model perfor-
mance. AutoML, for instance, exhibited a signicant decrease
of 0.113 in the R2 score, dropping from the original 0.971 to
0.858. Similar reductions were observed for LGB (0.154), XGB
(0.151), and RF (0.165), underlining the effectiveness of the
abnormal data removal process, albeit with a reduction in R2
scores.
Moving forward, the optimisation of DBSCAN parameters
led to additional improvements. AutoML recovered its R2
score by 0.003 compared to the original data, reaching 0.974.
Simultaneously, LGB, XGB, and RF experienced minor in-
creases of 0.006, 0.001, and 0.012, respectively, showcasing a
nuanced improvement over the original data. The subsequent
TABLE 4Performance comparison of ML models in abnormal detection for Romgberg Test: VP0006 dataset.
Data Models MAE MSE RMSE R2 RMSLE MAPE
VP0006 LGB 0.044 0.003 0.055 0.961 0.0313 0.0763
XGB 0.046 0.0032 0.056 0.96 0.0317 0.0758
RF 0.048 0.0034 0.058 0.959 0.032 0.0747
AutoML 0.042 0.0031 0.054 0.962 0.0315 0.0772
VP0006 (DBSCAN paper parameters) LGB 0.056 0.0046 0.067 0.931 0.036 0.085
XGB 0.058 0.0049 0.070 0.928 0.0365 0.087
RF 0.060 0.0051 0.071 0.925 0.037 0.089
AutoML 0.054 0.0043 0.065 0.933 0.035 0.084
VP0006 (DBSCAN optimise parameters) LGB 0.018 0.0012 0.034 0.962 0.020 0.065
XGB 0.017 0.0010 0.031 0.970 0.019 0.062
RF 0.019 0.0013 0.036 0.961 0.021 0.068
AutoML 0.016 0.0010 0.032 0.974 0.018 0.060
VP0006 hybrid (optimal DBSCAN þIQR) LGB 0.019 0.0013 0.036 0.971 0.016 0.055
XGB 0.018 0.0012 0.034 0.987 0.015 0.053
RF 0.018 0.0011 0.033 0.985 0.014 0.052
AutoML 0.016 0.0010 0.032 0.999 0.014 0.050
Note: Bold values in the results Tables showing the Proposed model highest results.
Abbreviations: AutoML, automated machine learning; BBS, Berg Balance Scale; DBSCAN, density‐based spatial clustering of applications with noise; IQR, interquartile range; LGB, light
GB; MAE, mean absolute error; MAPE, mean absolute percentage error; MSE, mean squared error; RF, random forest; RMSE, root mean squared error; RMSLE, root mean squared
logarithmic error; XGB, XGBoost.
KHAN ET AL.
-
17
integration of the proposed optimised DBSCAN with IQR
marked a pivotal phase. AutoML showcased a substantial
improvement of 0.027 in the R2 score compared to the original
data, reaching an impressive 0.998. LGB, XGB, and RF also
exhibited improvements of 0.017, 0.013, and 0.012, respec-
tively, highlighting the cumulative impact of the integrated
approach over the baseline.
In quantifying the overall improvement from the original
Borg RPE data to the proposed model, AutoML demonstrated
a cumulative improvement of 0.027. At the same time, LGB,
XGB, and RF showcased aggregated improvements of 0.017,
0.013, and 0.012, respectively. These ndings afrm the
effectiveness of the comprehensive strategy in enhancing
model robustness and predictive accuracy, recognising the
necessary trade‐off in R2 scores during the Borg RPE data
renement process.
Figure 14b shows the results for TUG data; in the initial
assessment using models LGB, XGB, RF, and AutoML, we
established a baseline with R2 scores of 0.95, 0.952, 0.948, and
0.948, respectively. Implementing DBSCAN with paper pa-
rameters for abnormal data removal resulted in notable
changes. AutoML showed a signicant improvement of 0.006
in R2 score from the original 0.948, highlighting its resilience
to the impact of abnormal data. In contrast, other models,
including LGB (0.025), XGB (0.024), and RF (0.026), experi-
enced decreases in their R2 scores, indicating the trade‐off
between data cleansing and the predictive performance of
these models. Furthermore, the optimisation of DBSCAN
parameters led to partial enhancements. AutoML demon-
strated a commendable increase of 0.010 in the R2 score
compared to the original data, reaching 0.958. Alternatively,
LGB, XGB, and RF experienced decreases of 0.009, 0.001, and
0.012, respectively. These observations highlight the rene-
ment process's impact on these models' predictive perfor-
mance. The subsequent integration of the proposed optimised
DBSCAN with IQR marked a pivotal phase. AutoML
exhibited a noteworthy increase of 0.045 in the R2 score
compared to the original data, reaching an impressive 0.993.
LGB, XGB, and RF also demonstrated 0.013, 0.023, and 0.013
improvements, respectively. In quantifying the overall
improvement from the original TUG Data to the proposed
model, AutoML demonstrated a cumulative improvement of
0.045. At the same time, LGB, XGB, and RF showcased
aggregated improvements of 0.013, 0.023, and 0.013, respec-
tively. These results underscore the efcacy of the approach in
enhancing model robustness for TUG Data, recognising the
necessary trade‐off in R2 scores during the data renement
process.
The ndings in Figure 14c underscore the marked
improvement in model performance resulting from integrating
the innovative abnormal data detection approach for Romberg
test data. The models—LGB, XGB, RF, and AutoML—initially
exhibited R2 scores of 0.961, 0.96, 0.959, and 0.962, respec-
tively. Introducing DBSCAN with Ref. [61] parameters led to a
substantial decrease in model performance. However, we
observed a noteworthy enhancement in the ne‐tuned
TABLE 5Performance comparison of ML models in abnormal detection for BBS: VP0007 dataset.
Data Models MAE MSE RMSE R2 RMSLE MAPE
VP0007 LGB 0.061 0.0045 0.0671 0.935 0.040 0.085
XGB 0.055 0.004 0.0632 0.915 0.038 0.082
RF 0.049 0.0037 0.0608 0.925 0.036 0.079
AutoML 0.043 0.0033 0.0575 0.942 0.034 0.076
VP0007 (DBSCAN paper parameters) LGB 0.028 0.0022 0.047 0.931 0.026 0.072
XGB 0.027 0.0021 0.046 0.936 0.025 0.071
RF 0.029 0.0023 0.048 0.932 0.027 0.074
AutoML 0.025 0.0018 0.042 0.954 0.023 0.068
VP0007 (DBSCAN optimise parameters) LGB 0.022 0.0018 0.043 0.952 0.021 0.065
XGB 0.021 0.0016 0.040 0.959 0.020 0.062
RF 0.024 0.0020 0.045 0.950 0.022 0.068
AutoML 0.019 0.0014 0.037 0.968 0.019 0.059
VP0007 hybrid (optimal DBSCAN þIQR) LGB 0.017 0.0011 0.033 0.970 0.015 0.052
XGB 0.017 0.0010 0.032 0.985 0.014 0.050
RF 0.018 0.0012 0.034 0.975 0.016 0.053
AutoML 0.015 0.0009 0.030 0.997 0.013 0.048
Note: Bold values in the results Tables showing the Proposed model highest results.
Abbreviations: AutoML, automated machine learning; BBS, Berg Balance Scale; DBSCAN, density‐based spatial clustering of applications with noise; IQR, interquartile range; LGB, light
GB; MAE, mean absolute error; MAPE, mean absolute percentage error; MSE, mean squared error; RF, random forest; RMSE, root mean squared error; RMSLE, root mean squared
logarithmic error; XGB, XGBoost.
18
-
KHAN ET AL.
parameters of the optimised DBSCAN approach. Specically,
R2 scores increased marginally by 0.001 (LGB), 0.01 (XGB),
0.002 (RF), and 0.012 (AutoML) when compared to the orig-
inal results. This highlights the precision and effectiveness of
the abnormal data detection process in rening model accu-
racy. We further rened the proposed DBSCAN with IQR for
abnormal data detection in the Romberg test data. The results
were striking, particularly for AutoML, which exhibited a
remarkable increase of 0.037 in R2 score compared to the
original data, reaching an impressive 0.999. Notably, other
models also demonstrated improved performance upon inte-
grating the abnormal data detection technique, solidifying the
outstanding efcacy of the approach in elevating model
robustness for Romberg test data.
In Figure 14d, we examine the impact of the abnormal data
detection approach on predictive models for BBS data. Initial
R2 scores for LGB, XGB, RF, and AutoML were 0.935, 0.915,
0.925, and 0.942, respectively. Applying DBSCAN with paper
parameters led to nuanced shifts. LGB slightly decreased to
0.931 (−0.004), XGB improved to 0.936 (þ0.021), RF gained
to 0.932 (þ0.007), and AutoML increased signicantly to 0.954
(þ0.012). These changes highlight the impact of data cleansing
with DBSCAN. Fine‐tuning DBSCAN parameters consistently
improved scores. LGB increased to 0.952 (þ0.017), XGB to
0.959 (þ0.044), RF to 0.95 (þ0.025), and AutoML to 0.968
(þ0.026). Optimisation proved effective in enhancing predic-
tive accuracy. Integrating optimised DBSCAN with IQR
further improved results. LGB reached 0.97 (þ0.035), XGB
soared to 0.985 (þ0.07), RF improved to 0.975 (þ0.05), and
AutoML excelled at 0.997 (þ0.055). From the original BBS
data to the proposed model, AutoML showed an outstanding
cumulative improvement of 0.055, whereas LGB, XGB, and
RF exhibited aggregated improvements of 0.035, 0.07, and
0.05, respectively. These ndings underscore the potency of
the abnormal data detection strategy in augmenting model
robustness for BBS data, acknowledging the R2 score trade‐off
during data renement.
7.2
|
Performance analysis in terms of
MAPE
Shifting the focus to the MAPE, we delve into a comprehen-
sive assessment of model robustness beyond R2 scores. This
analysis provides additional insights into the efcacy of the
abnormal data detection approach across diverse clinical
datasets. The following comparisons in Figure 15 underscore
the impact of the methodology on MAPE, shedding light on
FIGURE 14 Four clinical data R2 score analyses in different abnormal detection scenarios. AutoML, automated machine learning; BBS, Berg Balance Scale;
DBSCAN, density‐based spatial clustering of applications with noise; IQR, interquartile range; RF, random forest.
KHAN ET AL.
-
19
the broader implications of the research in enhancing predic-
tive accuracy and reliability.
Figure 15a shows the original Borg RPE data MAPE values
for predictive models—LGB, XGB, RF, and AutoML—stood
at 0.0778, 0.0742, 0.0759, and 0.0657, respectively. When
applying DBSCAN with paper parameters for abnormal data
detection, there was a noticeable increase in errors across all
models: LGB rose to 0.1 (þ0.022), XGB increased to 0.093
(þ0.018), RF elevated to 0.105 (þ0.029), and AutoML grew to
0.085 (þ0.019). Subsequent ne‐tuning with optimised
DBSCAN parameters led to consistent decreases in MAPE
values compared to the original data: LGB decreased to 0.07
(−0.007), XGB decreased to 0.067 (−0.007), RF decreased to
0.072 (−0.004), and AutoML decreased to 0.065 (−0.001). The
integration of the proposed optimised DBSCAN with IQR
showcased further substantial reductions in MAPE values:
LGB decreased to 0.057 (−0.021), XGB decreased to 0.055
(−0.019), RF decreased to 0.06 (−0.016), and AutoML
decreased to 0.052 (−0.014). This highlights the consistent
improvement in model accuracy and error reduction achieved
through the abnormal data detection strategy for Borg RPE.
In the context of the TUG data presented in Figure 15b,
the discussion begins by examining the original MAPE values
for the prediction models—LGB, XGB, RF, and AutoML—
standing at 0.0521, 0.068, 0.0598, and 0.0519, respectively.
These initial values serve as a baseline for assessing the impact
of subsequent abnormal data detection techniques. Upon
implementing DBSCAN with paper parameters for abnormal
data removal, we observed noticeable shifts in model errors.
LGB experienced a signicant increase, reaching a MAPE of
0.09 (þ0.0379), whereas XGB, RF, and AutoML also demon-
strated rises to 0.085 (þ0.017), 0.095 (þ0.0352), and 0.075
(þ0.0231), respectively. These alterations highlight the trade‐
off between data cleansing and predictive accuracy. Fine‐
tuning the abnormal data detection process with optimised
DBSCAN parameters led to improvements in some models.
However, the MAPE values for LGB, XGB, RF, and AutoML
remained at 0.075, 0.07, 0.08, and 0.065, respectively. Notably,
the optimised DBSCAN approach exhibited a slight increase in
error for LGB (þ0.0229) and RF (þ0.0202) compared to the
original data. Subsequent renement through integrating the
proposed optimised DBSCAN with IQR reduced errors across
all models. LGB, XGB, RF, and AutoML demonstrated
decreased MAPE values, reaching 0.055 (−0.0266), 0.052
(−0.016), 0.057 (−0.0038), and 0.05 (−0.0019), respectively,
compared to the original data. These reductions underscore the
effectiveness of the comprehensive approach in enhancing
predictive accuracy for TUG data.
FIGURE 15 Analysing mean absolute percentage error errors in clinical data prediction in different abnormal detection mechanisms. AutoML, automated
machine learning; BBS, Berg Balance Scale; DBSCAN, density‐based spatial clustering of applications with noise; IQR, interquartile range; RF, random forest.
20
-
KHAN ET AL.
Furthermore, for analysing the Romberg Test Data, illus-
trated in Figure 15c, the original MAPE values for prediction
models—LGB, XGB, RF, and AutoML—stood at 0.0763,
0.0758, 0.0747, and 0.0772, respectively. The implementation
of DBSCAN with paper parameters resulted in increased errors
for all models: LGB (0.085, þ0.0087), XGB (0.087, þ0.0112),
RF (0.089, þ0.0143), and AutoML (0.084, þ0.0068). However,
rening the abnormal data detection process with optimised
DBSCAN parameters led to error reductions, yielding MAPE
values of 0.065, 0.062, 0.068, and 0.06, respectively, showcasing
the effectiveness of the optimisation. The subsequent inte-
gration of the proposed optimised DBSCAN with IQR marked
a notable phase, further decreasing errors for LGB (0.055,
−0.0213), XGB (0.053, −0.0228), RF (0.052, −0.0227), and
AutoML (0.05, −0.0272). These reductions highlight the suc-
cess of the integrated approach in rening predictive accuracy
for Romberg Test Data, emphasising the careful trade‐offs
inherent in the abnormal data detection process.
Moreover, by examining the BBS Data, represented in
Figure 15d, the original MAPE values for prediction models—
LGB, XGB, RF, and AutoML—were 0.085, 0.082, 0.079, and
0.076, respectively. The application of DBSCAN with paper
parameters led to notable changes in errors, resulting in MAPE
values of 0.072 (LGB, −0.013), 0.071 (XGB, −0.011), 0.074
(RF, −0.005), and 0.068 (AutoML, −0.008). Subsequent ne‐
tuning of the abnormal data detection process with opti-
mised DBSCAN parameters showed further improvements,
yielding MAPE values of 0.065 (LGB, −0.02), 0.062 (XGB,
−0.02), 0.068 (RF, −0.011), and 0.059 (AutoML, −0.017).
Integrating the proposed optimised DBSCAN with IQR
marked a signicant phase, resulting in decreased errors for
LGB (0.052, −0.033), XGB (0.05, −0.032), RF (0.053, −0.026),
and AutoML (0.048, −0.028). These reductions underscore the
effectiveness of the integrated approach in rening predictive
accuracy for BBS Data, highlighting the nuanced trade‐offs
involved in the abnormal data detection process.
7.3
|
Comparison with existing studies
The comparison table, Table 6, provides a concise overview of
the performance improvements achieved through various data
preprocessing and abnormal data detection techniques across
different state‐of‐the‐art studies. Notably, the studies by Riyaz
et al. [62], Praveen et al. [63], and Ripan et al. [65] applied
standard statistical and clustering techniques, such as weighted
averages, IQR, and K‐means clustering, achieving modest
improvements in their respective datasets. These improve-
ments ranged from 2.64% to 3.77%, demonstrating a signi-
cant enhancement in model accuracy and robustness due to
data preprocessing. Similarly, methods such as principal
component analysis (PCA) and Z‐Score combined with IQR,
as used by Ozsahin et al. [64], and Kukkala et al. [66], showed
notable but slightly smaller gains, particularly on datasets such
as Wisconsin Breast Cancer and Pima Indian, where PCA
yielded improvements of 1.5% and 1%, respectively. These
outcomes underscore the value of traditional preprocessing
methods in improving model accuracy by reducing noise, albeit
with limitations in handling complex data abnormalities. In
contrast, the proposed study demonstrates signicantly higher
performance improvements across diverse clinical datasets
using an optimised DBSCAN combined with IQR. As high-
lighted in Table 6, the proposed approach resulted in cumu-
lative improvements of 2.67%, 4.5%, 3.7%, and 5.5% for Borg
RPE, TUG, Romberg Test, and BBS datasets, respectively.
These improvements exceed those achieved in prior studies,
especially for data like TUG and BBS, where model robustness
was substantially elevated. This enhanced performance is
attributed to the proposed abnormal data detection method-
ology, which effectively identies and removes noise and
outliers. By rening DBSCAN parameters and integrating it
with IQR, the proposed approach addresses the limitations of
conventional techniques. It optimally enhances predictive ac-
curacy across complex clinical datasets, positioning it as a su-
perior method in abnormal data preprocessing and model
performance enhancement.
8
|
CONCLUSION
Due to the dynamic and immeasurable characteristics,
abnormal detection in the clinical domain faces diverse chal-
lenges. The most prominent issue is detecting abnormality in
big data without compromising data quality or originality.
TABLE 6Comparison of the performance improvements for various models and datasets.
Paper Model Data Performance improvement (%)
Riyaz et al. [62] Weighted average Heart disease data 2.70
Praveen et al. [63] IQR Diabetes dataset 2.64
Ozsahin et al. [64] PCA Wisconsin breast and heart dataset and Pima Indian dataset 1.5 and 1
Ripan et al. [65] K‐means clustering Heart disease data 3.77
Kukkala et al. [66] Z‐score and IQR Heart disease data 3.0
Uddin et al. [67] IF and KDE WQ (EPA database) 3.0
Proposed study Optimal DBSCAN þIQR Borg RPE, TUG, Romgberg test, and BBS datasets 2.67, 4.5, 3.7, and 5.5
Abbreviations: BBS, Berg Balance Scale; DBSCAN, density‐based spatial clustering of applications with noise; IQR, interquartile range; PCA, principal component analysis; RPE, rating
of perceived exertion.
KHAN ET AL.
-
21
Therefore, this study established a robust hybrid scheme for
detecting anomalous data, combining heuristic and stochastic
methods to adeptly manage the complexities inherent in clin-
ical data. Initially, an optimal DBSCAN technique served as the
heuristic component to cluster the clinical data of the patients.
Subsequently, the IQR method was implemented to identify
anomalies in data patterns. This method leverages the differ-
ence between the rst quartile (Q1) and the third quartile (Q3)
to detect irregularities. Furthermore, sophisticated regression
models employing AutoML were constructed to evaluate the
inuence of this hybrid approach on detecting abnormal pat-
terns. Various evaluative analyses were also undertaken to
ascertain the effectiveness and signicance of this research
investigation.
Analysis from the experiments shows that the optimised
regression model, enhanced by AutoML within the proposed
hybrid anomaly detection framework, attained an R2 and
MAPE of 0.998 and 0.052 for the data VP0004 predictions.
Likewise, the model scored an R2 of 0.9993 and MAPE of
0.050 for VP0005, an R2 of 0.999 and MAPE of 0.050 for
VP0006, and an R2 of 0.997 and MAPE of 0.048 for VP0007.
Moreover, there was an enhancement in the R2 scores for
cutting‐edge models such as LGB, XGB, and RF within the
hybrid framework. Specically, R2 scores increased by 1.7%,
1.3%, and 1.2%, respectively, for VP0004 predictions. For
VP0005, the improvements were 1.3%, 2.3%, and 1.3%
respectively. For VP0006, LGB's performance increased by
1.0%, XGB by 2.7%, and RF by 2.6%. Similar enhancements
were observed with VP0007 data. This effectiveness analysis
demonstrates that the performance of the predictive models
was notably advanced by implementing the hybrid anomaly
detection strategy, which has proven reliable and effective in
managing deviations in clinical data. This ensures only perti-
nent information is relayed to healthcare professionals, aiding
in more accurate and efcient decision‐making. Looking ahead,
achieving interpretability remains crucial within healthcare
contexts. Moving forward, future efforts will focus on devel-
oping a hierarchical rules‐based model to improve the trans-
parency and adaptability of the hybrid approach. Additionally,
we plan to expand this hybrid method to include multivariate
time series data. Additionally, we will delve into developing
deep learning‐enriched regression models to comprehensively
assess the efcacy and pertinence of the hybrid model in
detecting abnormal patterns within patient clinical data.
ACKNOWLEDGEMENTS
This work was supported by the National Research Foundation
of Korea (NRF) grant funded by the Korean government
(MSIT) (No. RS‐2024‐00346238) for the project titled ‘A Study
of Decentralized Federated Learning Optimization Technolo-
gies for Autonomous On‐Device AI Networks’. It was also
supported by the NRF grant (No. RS‐2024‐00423362), funded
by the Korean government (MSIT). Any correspondence
related to this paper should be addressed to Do‐Hyeun Kim.
This work was supported by the 2024 education, research and
student guidance grant funded by Jeju National University.
CONFLICT OF INTEREST STATEMENT
The authors declare no conicts of interest.
DATA AVAILABILITY STATEMENT
The data that support the ndings of this study are available
from the corresponding author upon reasonable request.
ORCID
Murad Ali Khan
https://orcid.org/0009-0005-4865-7065
Harun Jamil https://orcid.org/0009-0001-8685-4407
REFERENCES
1. Kim, B.J., Tomprou, M.: The effect of healthcare data analytics training
on knowledge management: a quasi‐experimental eld study. J Open
Innov. Technol. Mark. Complex. 7(1), 60 (2021). https://doi.org/10.
3390/joitmc7010060
2. Shrestha, K., Naysary, B., Philip, S.S.S.: Fintech market efciency: a
multifractal detrended uctuation analysis. Finance Res. Lett. 54, 103775
(2023). https://doi.org/10.1016/j.frl.2023.103775
3. Iqbal, N., et al.: Enhanced time‐constraint aware tasks scheduling
mechanism based on predictive optimization for efcient load balancing
in smart manufacturing. J. Manuf. Syst. 64, 19–39 (2022). https://doi.
org/10.1016/j.jmsy.2022.05.015
4. Galetsi, P., Katsaliaki, K.: A review of the literature on big data analytics
in healthcare. J. Oper. Res. Soc. 71(10), 1511–1529 (2020). https://doi.
org/10.1080/01605682.2019.1630328
5. Khraisat, A., et al.: Survey of intrusion detection systems: techniques,
datasets and challenges. Cybersecurity 2(1), 1–22 (2019). https://doi.org/
10.1186/s42400‐019‐0038‐7
6. Goldstein, M., Uchida, S.: A comparative evaluation of unsupervised
anomaly detection algorithms for multivariate data. PLoS One 11(4),
e0152173 (2016). https://doi.org/10.1371/journal.pone.0152173
7. Dean, D.A., et al.: Scaling up scientic discovery in sleep medicine: the
national sleep research resource. Sleep 39(5), 1151–1164 (2016). https://
doi.org/10.5665/sleep.5774
8. Friedrich, B., Sawabe, T., Hein, A.: Unsupervised statistical concept drift
detection for behaviour abnormality detection. Appl. Intell. 53(3), 2527–
2537 (2023). https://doi.org/10.1007/s10489‐022‐03611‐3
9. Pang, G., Ting, K.M., Albrecht, D.: LeSiNN: detecting anomalies by
identifying least similar nearest neighbours. In: 2015 IEEE International
Conference on Data Mining Workshop (ICDMW), pp. 623–630. IEEE
(2015)
10. Khan, M.A., et al.: An optimized ensemble prediction model using
autoML based on soft voting classier for network intrusion detection. J.
Netw. Comput. Appl. 212, 103560 (2023). https://doi.org/10.1016/j.
jnca.2022.103560
11. Zhou, C., Paffenroth, R.C.: Anomaly detection with robust deep autoen-
coders. In: Proceedings of the 23rd ACM SIGKDD International Con-
ference on Knowledge Discovery and Data Mining, pp. 665–674 (2017)
12. Du, H., et al.: Novel clustering‐based approach for local outlier detection.
In: 2016 IEEE Conference on Computer Communications Workshops
(INFOCOM WKSHPS), pp. 802–811. IEEE (2016)
13. Metcalfe, A.V., Cowpertwait, P.S.: Introductory Time Series with R.
Springer (2009)
14. Kumar, S.V., Vanajakshi, L.: Short‐term trafc ow prediction using
seasonal arima model with limited input data. Eur. Transp. Res. Rev. 7(3),
1–9 (2015). https://doi.org/10.1007/s12544‐015‐0170‐8
15. Wang, Q., Song, X., Li, R.: A novel hybridization of nonlinear grey model
and linear arima residual correction for forecasting us shale oil produc-
tion. Energy 165, 1320–1331 (2018). https://doi.org/10.1016/j.energy.
2018.10.032
16. Aggarwal, C.C.: Outlier Analysis, 2nd ed (2016)
17. Hadi, A.S., Imon, A.R., Werner, M.: Detection of outliers. Wiley Inter-
discip. Rev.: Comput. Stat. 1(1), 57–70 (2009). https://doi.org/10.1002/
wics.6
22
-
KHAN ET AL.
18. Li, Z., et al.: COPOD: copula‐based outlier detection. In: 2020 IEEE
International Conference on Data Mining (ICDM), pp. 1118–1123
(2020)
19. Penny, K.I., Jolliffe, I.T.: A comparison of multivariate outlier detection
methods for clinical laboratory safety data. J. Roy. Stat. Soc.: Series D
(The Statistician) 50(3), 295–307 (2001). https://doi.org/10.1111/1467‐
9884.00279
20. Gao, S., Li, G., Wang, D.: A new approach for detecting multivariate
outliers. Commun. Stat. Theor. Methods 34(8), 1857–1865 (2005).
https://doi.org/10.1081/sta‐200066315
21. Southworth, H.: Detecting outliers in multivariate laboratory data. J.
Biopharm. Stat. 18(6), 1178–1183 (2008). https://doi.org/10.1080/
10543400802369046
22. Dang, X., Sering, R.: Nonparametric depth‐based multivariate outlier
identiers, and masking robustness properties. J. Stat. Plann. Inference
140(1), 198–213 (2010). https://doi.org/10.1016/j.jspi.2009.07.004
23. Fernando, T., et al.: Deep learning for medical anomaly detection–a
survey. ACM Comput. Surv. 54(7), 1–37 (2021). https://doi.org/10.
1145/3464423
24. Behrad, F., Abadeh, M.S.: An overview of deep learning methods for
multimodal medical data mining. Expert Syst. Appl. 200, 117006 (2022).
https://doi.org/10.1016/j.eswa.2022.117006
25. Sun, L., et al.: FogMed: a fog‐based framework for disease prognosis
based medical sensor data streams. Comput. Mater. Continua 66(1), 603–
619 (2021). https://doi.org/10.32604/cmc.2020.012515
26. Khoei, T.T., et al.: Ensemble learning methods for anomaly intrusion
detection system in smart grid. In: 2021 IEEE International Conference
on Electro Information Technology (EIT), pp. 129–135. IEEE (2021)
27. Nizan, O., Tal, A.: K‐NNN: nearest neighbors of neighbors for anomaly
detection. arXiv Preprint arXiv:2305.17695 (2023)
28. Zheng, Y., et al.: D2a: a dataset built for AI‐based vulnerability detection
methods using differential analysis. In: 2021 IEEE/ACM 43rd Interna-
tional Conference on Software Engineering: Software Engineering in
Practice (ICSE‐SEIP), pp. 111–120. IEEE (2021)
29. Schlegl, T., et al.: f‐AnoGAN: fast unsupervised anomaly detection with
generative adversarial networks. Med. Image Anal. 54, 30–44 (2019).
https://doi.org/10.1016/j.media.2019.01.010
30. Shehata, M., et al.: Computer‐aided diagnostic system for early detection
of acute renal transplant rejection using diffusion‐weighted MRI. IEEE
Trans. Biomed. Eng. 66(2), 539–552 (2018). https://doi.org/10.1109/
tbme.2018.2849987
31. Fernando, T., et al.: Neural memory plasticity for medical anomaly
detection. Neural Netw. 127, 67–81 (2020). https://doi.org/10.1016/j.
neunet.2020.04.011
32. Thudumu, S., et al.: A comprehensive survey of anomaly detection
techniques for high dimensional big data. J. Big Data 7(1), 1–30 (2020).
https://doi.org/10.1186/s40537‐020‐00320‐x
33. Chalapathy, R., Chawla, S.: Deep learning for anomaly detection: a sur-
vey. arXiv Preprint arXiv:1901.03407 (2019)
34. Li, J., et al.: Outlier detection using iterative adaptive mini‐minimum
spanning tree generation with applications on medical data. Front.
Physiol. 14, 1233341 (2023). https://doi.org/10.3389/fphys.2023.
1233341
35. Li, J., et al.: MS2OD: outlier detection using minimum spanning tree and
medoid selection. Mach. Learn.: Sci. Technol. 5(1), 015025 (2024).
https://doi.org/10.1088/2632‐2153/ad2492
36. Gijzel, S.M., et al.: Resilience in clinical care: getting a grip on the re-
covery potential of older adults. J. Am. Geriatr. Soc. 67(12), 2650–2657
(2019). https://doi.org/10.1111/jgs.16149
37. Ji, H., et al.: Effects of exercise‐based cardiac rehabilitation in patients with
acute coronary syndrome: a meta‐analysis. Med. Sci. Monit.: Int. J. Exp.
Clin. Res. 25, 5015–5027 (2019). https://doi.org/10.12659/msm.917362
38. Gevaert, A.B., et al.: Towards a personalised approach in exercise‐based
cardiovascular rehabilitation: how can translational research help? A ‘call
to action’ from the section on secondary prevention and cardiac reha-
bilitation of the European Association of Preventive Cardiology. Eur. J.
Prev. Cardiol. 27(13), 1369–1385 (2020). https://doi.org/10.1177/
2047487319877716
39. Morris, N.R., et al.: Exercise‐based rehabilitation programmes for pul-
monary hypertension. Cochrane Database Syst. Rev. 3(3), CD011285
(2023). https://doi.org/10.1002/14651858.cd011285.pub3
40. Babaei, N., et al.: A systematic review of the use of commercial wearable
activity trackers for monitoring recovery in individuals undergoing total
hip replacement surgery. Cyborg. Bionic Syst. 2022, 9794641 (2022).
https://doi.org/10.34133/2022/9794641
41. Dibben, G., et al.: Exercise‐based cardiac rehabilitation for coronary
heart disease. Cochrane Database Syst. Rev. 11(11), CD001800 (2021)
42. Cheng, L., et al.: Short‐and long‐term effects of pulmonary rehabilitation
for idiopathic pulmonary brosis: a systematic review and meta‐analysis.
Clin. Rehabil. 32(10), 1299–1307 (2018). https://doi.org/10.1177/
0269215518779122
43. Gu, Y., et al.: MFGTN: a multi‐modal fast gated transformer for iden-
tifying single trawl marine shing vessel. Ocean. Eng. 303, 117711 (2024).
https://doi.org/10.1016/j.oceaneng.2024.117711
44. Ahmed, I., et al.: Effect of pulmonary rehabilitation approaches on
dyspnea, exercise capacity, fatigue, lung functions and quality of life in
patients with Covid‐19: a systematic review and meta‐analysis. Arch.
Phys. Med. Rehabil. 103(10), 2051–2062 (2022). https://doi.org/10.
1016/j.apmr.2022.06.007
45. Xu, X., Wei, Z.: Dynamic pickup and delivery problem with trans-
shipments and LIFO constraints. Comput. Ind. Eng. 175, 108835 (2023).
https://doi.org/10.1016/j.cie.2022.108835
46. Chetambath, R.: Pulmonary rehabilitation in chronic respiratory diseases.
J. Adv. Lung Health 2(1), 5–12 (2022). https://doi.org/10.4103/jalh.jalh_
9_21
47. Amin, R., et al.: Effect of a home‐based pulmonary rehabilitation pro-
gram on functional capacity and health‐related quality of life in people
with interstitial lung disease–a randomized controlled trial protocol.
Respir. Med. 201, 106927 (2022). https://doi.org/10.1016/j.rmed.2022.
106927
48. Rasheed, K., et al.: Machine learning for predicting epileptic seizures
using EEG signals: a review. IEEE Rev. Biomed. Eng. 14, 139–155
(2020). https://doi.org/10.1109/rbme.2020.3008792
49. Li, F., et al.: Classication of heart sounds using convolutional neural
network. Appl. Sci. 10(11), 3956 (2020). https://doi.org/10.3390/
app10113956
50. Lu, S., et al.: Surgical instrument posture estimation and tracking based
on LSTM. ICT Express 10(3), 465–471 (2024). https://doi.org/10.1016/
j.icte.2024.01.002
51. Yang, C., et al.: A dual‐domain diffusion model for sparse‐view CT
reconstruction. IEEE Signal Process. Lett. 31, 1279–1283 (2024).
https://doi.org/10.1109/lsp.2024.3392690
52. Zhang, L., Yang, H., Jiang, Z.: Imbalanced biomedical data classication
using self‐adaptive multilayer elm combined with dynamic GAN. Bio-
med. Eng. Online 17(1), 1–21 (2018). https://doi.org/10.1186/s12938‐
018‐0604‐3
53. Liu, H., et al.: Taxonomy and real‐time classication of artifacts
during biosignal acquisition: a starter study and dataset of ECG. IEEE
Sens. J. 24(6), 9162–9171 (2024). https://doi.org/10.1109/jsen.2024.
3356651
54. Zhao, X., et al.: Diabetes‐related avoidable hospitalisations and its rela-
tionship with primary healthcare resourcing in China: a cross‐sectional
study from Sichuan province. Health Soc. Care Community 30(4),
e1143–e1156 (2022). https://doi.org/10.1111/hsc.13522
55. Amores, J., Maes, P., Paradiso, J.: Bin‐ary: detecting the state of organic
trash to prevent insalubrity. In: Adjunct Proceedings of the 2015 ACM
International Joint Conference on Pervasive and Ubiquitous Computing
and Proceedings of the 2015 ACM International Symposium on Wear-
able Computers, pp. 313–316 (2015)
56. Luo, Y., Chen, D., Xing, X.‐L.: Comprehensive analyses revealed eight
immune related signatures correlated with aberrant methylations as
prognosis and diagnosis biomarkers for kidney renal papillary cell car-
cinoma. Clin. Genitourin. Cancer 21(5), 537–545 (2023). https://doi.
org/10.1016/j.clgc.2023.06.011
57. Louiset, R., et al.: UCSL: a machine learning expectation‐maximization
framework for unsupervised clustering driven by supervised learning.
KHAN ET AL.
-
23
In: Machine Learning and Knowledge Discovery in Databases. Research
Track: European Conference, ECML PKDD 2021, Bilbao, Spain,
September 13–17, 2021, Proceedings, Part I 21, pp. 755–771. Springer
(2021)
58. Li, Q., et al.: LI‐EMRSQL: linking information enhanced Text2SQL
parsing on complex electronic medical records. IEEE Trans. Reliab.
73(2), 1280–1290 (2023)
59. Amanullah, M.R., et al.: Prognostic implications of associated cardiac
abnormalities detected on echocardiography in patients with moderate
aortic stenosis. Cardiovasc. Imaging 14(9), 1724–1737 (2021). https://
doi.org/10.1016/j.jcmg.2021.04.009
60. Peng, Y., et al.: Failure to identify robust latent variables of positive or
negative valence processing across units of analysis. Biol. Psychiatry:
Cogn. Neurosci. Neuroimaging 6(5), 518–526 (2021). https://doi.org/
10.1016/j.bpsc.2020.12.005
61. Braei, M., Wagner, S.: Anomaly detection in univariate time‐series: a
survey on the state‐of‐the‐art. arXiv Preprint arXiv:2004.00433 (2020)
62. Riyaz, L., Butt, M.A., Zaman, M.: Improving coronary heart disease
prediction by outlier elimination. Appl. Comput. Sci. 18(1), 70–88 (2022).
https://doi.org/10.35784/acs‐2022‐6
63. Praveen, S.P., et al.: Outlier management and its impact on diabetes
prediction: a voting ensemble study. J. Intell. Syst. Internet Things 12(1),
08–19 (2024). https://doi.org/10.54216/jisiot.120101
64. Ozsahin, D., et al.: Impact of outliers and dimensionality reduction on the
performance of predictive models for medical disease diagnosis. In:
International Conference on Articial Intelligence in Everything (AIE),
Lefkosa, Cyprus (2022). [Paper presentation]. https://doi.org/10.1109/
AIE57029
65. Ripan, R.C., et al.: A data‐driven heart disease prediction model through
k‐means clustering‐based anomaly detection. SN Comput. Sci. 2(2), 112
(2021). https://doi.org/10.1007/s42979‐021‐00518‐7
66. Kukkala, V.R., et al.: A study on outlier detection and feature engineering
strategies in machine learning for heart disease prediction. Comput. Syst.
Sci. Eng. 48(5), 1085–1112 (2024). https://doi.org/10.32604/csse.2024.
053603
67. Uddin, M.G., et al.: Data‐driven evolution of water quality models: an in‐
depth investigation of innovative outlier detection approaches‐a case
study of Irish Water Quality Index (IEWQI) model. Water Res. 255,
121499 (2024). https://doi.org/10.1016/j.watres.2024.121499
How to cite this article: Khan, M.A., et al.: Enhancing
patient rehabilitation predictions with a hybrid anomaly
detection model: density‐based clustering and
interquartile range methods. CAAI Trans. Intell.
Technol. 1–24 (2025). https://doi.org/10.1049/cit2.
70000
24
-
KHAN ET AL.
