ArticlePDF Available

Data processing pipeline for cardiogenic shock prediction using machine learning

Frontiers
Frontiers in Cardiovascular Medicine
Authors:

Abstract and Figures

Introduction Recent advances in machine learning provide new possibilities to process and analyse observational patient data to predict patient outcomes. In this paper, we introduce a data processing pipeline for cardiogenic shock (CS) prediction from the MIMIC III database of intensive cardiac care unit patients with acute coronary syndrome. The ability to identify high-risk patients could possibly allow taking pre-emptive measures and thus prevent the development of CS. Methods We mainly focus on techniques for the imputation of missing data by generating a pipeline for imputation and comparing the performance of various multivariate imputation algorithms, including k-nearest neighbours, two singular value decomposition (SVD)—based methods, and Multiple Imputation by Chained Equations. After imputation, we select the final subjects and variables from the imputed dataset and showcase the performance of the gradient-boosted framework that uses a tree-based classifier for cardiogenic shock prediction. Results We achieved good classification performance thanks to data cleaning and imputation (cross-validated mean area under the curve 0.805) without hyperparameter optimization. Conclusion We believe our pre-processing pipeline would prove helpful also for other classification and regression experiments.
This content is subject to copyright.
EDITED BY
Benedikt Schrage,
University Medical Center Hamburg-Eppendorf,
Germany
REVIEWED BY
Stefania Sacchi,
San Raffaele Scientic Institute (IRCCS), Italy
Meraj Neyazi,
University Medical Center Hamburg-Eppendorf,
Germany
Kishore Surendra,
University Medical Center Hamburg-Eppendorf,
Germany
*CORRESPONDENCE
Branislav Bezak
branislav.bezak@premedix.org
SPECIALTY SECTION
This article was submitted to Heart Failure and
Transplantation, a section of the journal
Frontiers in Cardiovascular Medicine
RECEIVED 27 December 2022
ACCEPTED 07 March 2023
PUBLISHED 23 March 2023
CITATION
Jajcay N, Bezak B, Segev A, Matetzky S,
Jankova J, Spartalis M, El Tahlawi M, Guerra F,
Friebel J, Thevathasan T, Berta I, Pölzl L,
Nägele F, Pogran E, Cader FA, Jarakovic M,
Gollmann-Tepeköylü C, Kollarova M,
Petrikova K, Tica O, Krychtiuk AA, Tavazzi G,
Skurk C, Huber K and Böhm A (2023) Data
processing pipeline for cardiogenic shock
prediction using machine learning.
Front. Cardiovasc. Med. 10:1132680.
doi: 10.3389/fcvm.2023.1132680
COPYRIGHT
© 2023 Jajcay, Bezak, Segev, Matetzky,
Jankova, Spartalis, El Tahlawi, Guerra, Friebel,
Thevathasan, Berta, Pölzl, Nägele, Pogran,
Cader, Jarakovic, Gollmann-Tepeköylü,
Kollarova, Petrikova, Tica, Krychtiuk, Tavazzi,
Skurk, Huber and Böhm. This is an open-access
article distributed under the terms of the
Creative Commons Attribution License (CC BY).
The use, distribution or reproduction in other
forums is permitted, provided the original
author(s) and the copyright owner(s) are
credited and that the original publication in this
journal is cited, in accordance with accepted
academic practice. No use, distribution or
reproduction is permitted which does not
comply with these terms.
Data processing pipeline for
cardiogenic shock prediction
using machine learning
Nikola Jajcay1,2, Branislav Bezak1,3,4*, Amitai Segev5,6,
Shlomi Matetzky5,6, Jana Jankova1, Michael Spartalis7,8,
Mohammad El Tahlawi9, Federico Guerra10, Julian Friebel11,
Tharusan Thevathasan11,12,13,14, Imrich Berta1, Leo Pölzl15,
Felix Nägele15, Edita Pogran16, F. Aaysha Cader17,
Milana Jarakovic18,19, Can Gollmann-Tepeköylü15, Marta Kollarova1,
Katarina Petrikova1, Otilia Tica20,21, Konstantin A. Krychtiuk22,23,
Guido Tavazzi24,25, Carsten Skurk11,13, Kurt Huber16
and Allan Böhm1,4,26
1
Premedix Academy, Bratislava, Slovakia,
2
Department of Complex Systems, Institute of Computer
Science, Czech Academy of Sciences, Prague, Czech Republic,
3
Clinic of Cardiac Surgery, National
Institute of Cardiovascular Diseases, Bratislava, Slovakia,
4
Faculty of Medicine, Comenius University in
Bratislava, Bratislava, Slovakia,
5
The Leviev Cardiothoracic & Vascular Center, Chaim Sheba Medical
Center, Ramat Gan, Israel,
6
Afliated to the Sackler Faculty of Medicine, Tel Aviv University, Tel Aviv, Israel,
7
3rd Department of Cardiology, National and Kapodistrian University of Athens, Athens, Greece,
8
Global
Clinical Scholars Research Training (GCSRT) Program, Harvard Medical School, Boston, MA, United States,
9
Department of Cardiology, Faculty of Human Medicine, Zagazig University, Zagazig, Egypt,
10
Cardiology
and Arrhythmology Clinic, Marche Polytechnic University, University Hospital Umberto I - Lancisi -
Salesi, Ancona, Italy,
11
Department of Cardiology Angiology and Intensive Care Medicine, Deutsches
Herzzentrum der Charité (DHZC), Campus Benjamin Franklin, Charité - Universitätsmedizin Berlin, Berlin,
Germany,
12
Berlin Institute of Health, CharitéUniversitätsmedizin Berlin, Berlin, Germany,
13
Deutsches
Zentrum für Herz-Kreislauf-Forschung e.V., Berlin, Germany,
14
Institute of Medical Informatics, Charité
Universitätsmedizin Berlin, Berlin, Germany,
15
Department for Cardiac Surgery, Cardiac Regeneration
Research, Medical University of Innsbruck, Innsbruck, Austria,
16
3rd Medical Department, Cardiology and
Intensive Care Medicine, Wilhelminen Hospital, Vienna, Austria,
17
Department of Cardiology, Ibrahim
Cardiac Hospital & Research Institute, Dhaka, Bangladesh,
18
Cardiac Intensive Care Unit, Institute for
Cardiovascular Diseases of Vojvodina, Sremska Kamenica, Serbia,
19
Faculty of Medicine, University of Novi
Sad, Novi Sad, Serbia,
20
Cardiology Department, Emergency County Clinical Hospital of Oradea, Oradea,
Romania,
21
Institute of Cardiovascular Sciences, University of Birmingham, Medical School, Birmingham,
United Kingdom,
22
Department of Internal Medicine II, Division of Cardiology, Medical University of
Vienna, Vienna, Austria,
23
Duke Clinical Research Institute Durham, NC, United States,
24
Department of
Clinical-Surgical, Diagnostic and Paediatric Sciences, University of Pavia, Pavia, Italy,
25
Anesthesia and
Intensive Care, Fondazione Policlinico San Matteo Hospital IRCCS, Pavia, Italy,
26
Department of Acute
Cardiology, National Institute of Cardiovascular Diseases, Bratislava, Slovakia
Introduction: Recent advances in machine learning provide new possibilities to
process and analyse observational patient data to predict patient outcomes. In this
paper, we introduce a data processing pipeline for cardiogenic shock (CS)
prediction from the MIMIC III database of intensive cardiac care unit patients with
acute coronary syndrome. The ability to identify high-risk patients could possibly
allow taking pre-emptive measures and thus prevent the development of CS.
Methods: We mainly focus on techniques for the imputation of missing data by
generating a pipeline for imputation and comparing the performance of various
multivariate imputation algorithms, including k-nearest neighbours, two singular
value decomposition (SVD)based methods, and Multiple Imputation by Chained
Equations. After imputation, we select the nal subjects and variables from the
imputed dataset and showcase the performance of the gradient-boosted
framework that uses a tree-based classier for cardiogenic shock prediction.
Results: We achieved good classication performance thanks to data cleaning and
imputation (cross-validated mean area under the curve 0.805) without
hyperparameter optimization.
TYPE Original Research
PUBLISHED 23 March 2023
|
DOI 10.3389/fcvm.2023.1132680
Frontiers in Cardiovascular Medicine 01 frontiersin.org
Conclusion: We believe our pre-processing pipeline would prove helpful also for other
classication and regression experiments.
KEYWORDS
classication, machine learning, missing data imputation, processing pipeline, prediction model,
cardiogenic shock
1. Introduction
Modern technology, increasing computing power, and
advances in machine learning provide new possibilities to process
and extract maximum knowledge from available patient data that
can improve healthcare, patient outcomes (14), and new
frontiers in predictive medicine (5).
The MIMIC dataset (Medical Information Mart for Intensive
Care) (6) is a widely-used public data source including over fty
thousand de-identied electronic health records (EHR) of patients
admitted to critical care units at Beth Israel Deaconess Medical
Center in Boston, MA, the USA, from 2001 to 2012. This database
contains a large amount of clinical data, which resulted in several
analytic studies in cardiovascular medicine (79).
Unfortunately, the data analysis requires a cautious pre-analytic
phase of meticulous data cleaning and processing which may be
particularly challenging in multi-national observational studies and
registries (10). In this paper, we describe in detail our methodology
for processing the MIMIC dataset as a part of developing a scoring
system for predicting cardiogenic shock (CS) in patients suffering
from acute coronary syndrome (ACS) (11).
Despite improvements in diagnostic and treatment options, CS
still affects 10% of ACS patients with unacceptably high, reaching
nearly 50% mortality (12). CS is not only an isolated decrease in
cardiac function but a rapidly progressing multiorgan
dysfunction accompanied by severe cellular and metabolic
abnormalities, and when developed, even the elimination of the
underlying primary cause is not to reverse this vicious circle
(13). The aim of the STOPSHOCK project is to derivate and
validate a simple scoring system able to identify high-risk
patients prior to the development of CS. Such patient
stratication could allow us to take pre-emptive measures, such
as the implantation of percutaneous mechanical circulatory
support, and thus prevent the development of CS, ultimately
leading to improved survival of ACS patients.
2. Methods
2.1. First cohort selection
In our study, we included patients presenting with ACS
undergoing cardiac catheterization. The cohort was then divided
into two groups: a patient group, comprising patients who
developed cardiogenic shock during hospitalization, and a control
group, comprising patients who did not develop cardiogenic shock.
The patients were selected and assigned to a cohort based on the
diagnosis and procedures undertaken during the hospitalization.
The identication of diagnosis and management was made
using the ICD9 coding scheme (14).
The ICD9 codes for both cohorts are detailed in Table 1.
Briey, the control group included patients with acute
myocardial infarction, ischemic heart disease, and angina pectoris
undertaking cardiac catheterization, but excluded cardiogenic or
unspecied shock. Conversely, the patient group contained
patients who developed cardiogenic shock, in addition to the
myocardial infarction diagnoses and catheterization codes.
The nal number of unique hospital stays for our control group
reached 3,056, while we included 703 hospital stays for the patient
group.
However, based on coded data, it was not possible to reliably
distinguish between patients who were already admitted with
shock and those who developed shock during hospitalization.
Several methods were tested based on the variation of patient
variables (blood pressure, heart rate, use of inotropes, uid
TABLE 1 Initial cohort selection based on ICD9-coded diagnoses and
procedures.
ICD9
diagnosis
title patients controls
78551 Cardiogenic shock x
78550 Shock, unspecied x
4100041092 Various versions of Acute
myocardial infarction
xx
41189 Other acute and subacute forms of
ischemic heart disease, other
xx
4139 Other and unspecied angina
pectoris
xx
ICD9
procedures
title patients controls
0066 Percutaneous transluminal coronary
angioplasty (PTCA)
xx
3604 Intracoronary artery thrombolytic
infusion
xx
3606 Insertion of non-drug-eluting
coronary artery stent(s)
xx
3607 Insertion of drug-eluting coronary
artery stent(s)
xx
3609 Other removal of coronary artery
obstruction
xx
8855 Coronary arteriography using a
single catheter
xx
8856 Coronary arteriography using two
catheters
xx
8857 Other and unspecied coronary
arteriography
xx
3722 Left heart cardiac catheterization x x
3723 Combined right and left heart
cardiac catheterization
xx
total n. 703 3056
ICD, international classication of diseases.
Jajcay et al. 10.3389/fcvm.2023.1132680
Frontiers in Cardiovascular Medicine 02 frontiersin.org
replacement therapy, or similar) (15,16). However, none of these
methods provided reliable results when veried based on textual
hospitalization summaries.
Ultimately, we analyzed individual discharge reports provided
within the dataset. A total of 703 summaries were manually
sorted, which resulted in 172 unique hospital admissions of
patients who developed cardiogenic shock during the hospital
course.
2.2. Data inspection
All available data were inspected, plotted, and sorted based on
missing values. Potentially relevant clinical variables to the aim of
the study were selected. As several variables are stored in the
database using multiple codes for the same variable, the ones
selected were clustered into aggregated variables. For example,
systolic blood pressure is available as Non-Invasive Blood Pressure
systolic, Arterial BP [Systolic], Manual Blood Pressure Systolic
Left, Manual Blood Pressure Systolic Right, Arterial Blood
Pressure systolic, ART BP Systolic, Manual BP [Systolic]. In our
case this clustering concerned mean, systolic and diastolic arterial
pressures. For this scoring system, the rst recorded variables
were selected.
When different units of measure were used, they were
converted to the international standard. Outliers were inspected
manually. Some values were manually corrected (e.g., 375 °C to
37.5 °C). Extremal values above the common thresholdclearly
incorrectly entered values (e.g., the body temperature of 5 °C)
were deleted.
In the preselected 84 variables, 7.86% of missing values were
found (Figure 1) and the missing data were missing completely
at random (17).
2.3. Pre-imputation cohort and variables
selection
In order to improve the predictive ability of the scoring system,
missing values of preselected features were imputed. Many
different univariate (18,19) and multivariate techniques (18,20)
have been described for data imputation. We used multivariate
techniques considering the rate of missing values that were
missing completely at random and the relatively high number of
variables. Additionally, multivariate imputation techniques can
accommodate and mimic interdependencies between variables
(21), which seemed more appropriate for the current study. To
improve the predictive value of multivariate imputation
FIGURE 1
The percentage of missing values per variable in our selection of patients.
Jajcay et al. 10.3389/fcvm.2023.1132680
Frontiers in Cardiovascular Medicine 03 frontiersin.org
techniques, we decided to enlarge our initial groups, include more
variables, and expand our original dataset both in terms of patients
and variables included.
We merged patient and control groups into one cohort for the
imputation. Next, we included similar patients by reducing the
inclusion criteria to:
Patients with at least one ICD9 diagnosis code:
78551: Cardiogenic shock
78550: Shock, unspecied
4100041092: various versions of Acute myocardial
infarction
41189: Other acute and subacute forms of ischemic heart
disease, other
4139: Other and unspecied angina pectoris
The advantage of selecting patients with a related diagnosis
compared to random selection is a greater similarity of data,
which should theoretically result in higher imputation accuracy
(22).
As for additional variables used for the sole purpose of missing
data imputation, we included 19 clinically relevant additional
variables with minimal missing values. These 19 additional
variables are detailed in Table 2. The nal pre-imputation
dataset contained 4,595 patients in one grouped cohort and 86
variables.
2.4. Data imputation
We utilized Multiple Imputation by Chained Equations
(MICE) (2325) as our primary algorithm for the missing data
imputation. The MICE algorithm imputes missing data through
an iterative series of predictive models. In each iteration,
specied variables in the dataset are imputed using other
variables. These iterations are run until it appears that
convergence has been met. Gradient-boosted, tree-based
predictive models were implemented as a part of the LightGBM
package (26). Moreover, the predictive mean matching technique
(PMM) was also used during the imputation (27). PMM entails
the selection of a data point from the original, non-missing data
with a predicted value close to the predicted value of the missing
sample. The closest ve values are chosen as candidates, from
which a value is sampled randomly. By using PMM, we could
correctly impute variables with the multimodal empirical
distribution. By exploiting the stochastic nature of tree-based
predictive models, we could impute multiple versions of the
dataset. This allowed us to run a sensitivity analysis and assess
the effect of missing data on our nal classication model. As a
good balance between computational time and statistical power,
we decided to run the imputation ten times. In order to have a
benchmark for our stability analysis, we further selected three
additional imputation algorithms: k-Nearest Neighbors (KNN)
(28), Soft Impute(performs matrix completion by iterative soft
thresholding of SVD decomposition) (29), and Iterative SVD
(performs matrix completion by iterative low-rank SVD
decomposition) (30).
TABLE 2 Overview of the variables from the MIMIC III database that were
used as a pre-imputation dataset.
variable missing
values (%)
variable missing
values (%)
Pre-Admission
Intake
83.2 RBC 2.68
BSA 46.99 WBC (411,000) 2.55
Admit Hit 46.92 Braden Score 2.52
Pain Cause 33.97 Activity 2.46
Arterial BP [Systolic] 30.66 Hemoglobin 2.44
Arterial BP Mean 30.49 Orientation 2.33
Pain Management
[Route/Status #1]
30.1 Glasgow Coma Scale 2.29
Arterial BP
[Diastolic]
29.49 Heart Rhythm 2.24
Admit Wt 23.31 Platelets 2.15
age 23.5 Oral Cavity 2.15
sex 17.74 Bowel Sounds 2.11
Procedures 17.8 Sodium (135148) 2.0
Riker-SAS Scale 15.63 Abdominal Assessment 1.92
Calcium (8.410.2) 14.15 Level of Conscious 1.89
Phosphorous (2.74.5) 13.6 BUN (620) 1.87
Marital Status 13.45 Glucose (70105) 1.87
O2 Flow (lpm) 12.38 RLL Lung Sounds 1.85
Service Type 11.19 RUL Lung Sounds 1.85
PT (1113.5) 9.42 LUL Lung Sounds 1.83
INR (24 ref. range) 9.23 Respiratory Pattern 1.83
PTT (2235) 9.8 Creatinine (01.3) 1.83
Religion 8.27 LLL Lung Sounds 1.81
Pain Location 8.3 Hematocrit 1.57
Temperature F 6.88 Potassium (3.55.3) 1.55
NBP [Systolic] 6.59 DG: Tobacco 0.0
Temperature C 6.46 DG: Pure
hypercholesterolem
0.0
NBP Mean 6.38 EKG0 0.0
Position 5.81 DG: Hypertension
NOS
0.0
Behavior 5.72 Dorsal PedPulse [Left] 0.0
PostTib. Pulses
[Right]
5.35 DG: Hyperlipidemia
NEC/NOS
0.0
NBP [Diastolic] 4.98 DG: Fam hx-ischem
heart dis
0.0
Magnesium (1.62.6) 4.48 DG: CHF NOS 0.0
PostTib. Pulses [Left] 4.31 EKG 0.0
Urine [Color] 3.58 Allergy 3 0.0
Urine [Appearance] 3.74 Ectopy Type 0.0
Activity Tolerance 3.55 Diet Type 0.0
SpO2 3.44 Dorsal PedPulse
[Right]
0.0
Assistance Device 3.33 Allergy 1 0.0
Pain Type 3.18 Ectopy Frequency 0.0
Heart Rate 2.89 Allergy 0.0
Chloride (100112) 2.76 O2 Delivery Device 0.0
Carbon Dioxide 2.72 Readmission 0.0
Respiratory Rate 2.7 Pain Present 0.0
The table shows the percentage of missing values per variable in our selection of
4595 patients.
Jajcay et al. 10.3389/fcvm.2023.1132680
Frontiers in Cardiovascular Medicine 04 frontiersin.org
2.5. Final cohort and variable selection
After successfully imputing all 13 datasets (10 with MICE, 1
with KNN, 1 with Soft Impute, and 1 with Iterative SVD), the
nal selection of the control and patient group, and variables to
be used for the diagnostic model was made. Here, we used our
initial selection and discarded added patients and variables.
2.6. Computational methods
All analyses were performed in python version 3.8.13 (https://
www.python.org) with appropriate packages (pandas 1.4.2, scipy
1.8.0, pingouin 0.5.1, miceforest 5.4.0, lightgbm 3.3.2, seaborn 0.11.
2). The repository containing the analysis code will be available
after the nalization of this study or upon reasonable request.
3. Results
3.1. Imputed nal dataset statistics
As the rststep,thepercentageofmissingvaluesof
preselected variables was plotted (cf. section Pre-imputation
cohort and variables selection). The results are shown in
Figure 1.
Evaluating the imputation quality is not straightforward, and
universally accepted pipelines do not exist (31). We opted for
visual assessment to qualitatively estimate the quality of
imputation and compare distributions of imputed data with
original, non-missing data employing the Kolmogorov-Smirnov
test (32) for distribution equivalent for the quantitative
assessment. The example of imputation quality for selected
variables is shown in Figure 2.
The MICE imputation algorithm correctly captured the data
distribution in most cases (Figure 2) including multimodal
distribution (e.g., O
2
ow) differently from other imputation
techniques, such as Iterative SVD, KNN, or Soft Impute.
The quality of imputation on the whole dataset was performed
by comparing the distributions demonstrating approximately 20
signicant differences between the original and imputed datasets.
At the same time, other methods exhibit almost twice as many
(Figure 3).
The right panel of Figure 3 shows the number of signicant
differences per variable in all imputation methods. Variables with
a high number are hard to impute. Naturally, this correlates
with the percentage of initially missing data (cf. Figure 1), and
categorical variables with many different categories (e.g., Lung
FIGURE 2
Violin plots showing distributions of initially non-missing data ( purple), imputed data (blue), and all data (red) for three imputation methods (with the MICE
method, we imputed ten datasets). The distributions are shown for a selection of variables: Heart rate, Respiratory rate, O
2
ow [lpm], Glucose, Arterial
Systolic Blood Pressure, and Riker-SAS scale.
Jajcay et al. 10.3389/fcvm.2023.1132680
Frontiers in Cardiovascular Medicine 05 frontiersin.org
Sounds, Hearth Rhythm, Respiratory Pattern, and others) are
harder to impute. The ability to correctly impute a variable, as
shown in Figure 3, will be considered for a nal variable selection.
Finally, we also visualized a detailed plot of the ability of our
imputation algorithms to estimate the variable distribution, as
shown in Figure 4. With very few exceptions, MICE-imputed
datasets generally show lower K-S statistics, therefore achieving a
better match between imputed and original distribution.
Overall, we observed the superior performance of the MICE
method, as for most of the variables, it provides distributions of
imputed values closer to the original (i.e., lower Kolmogorov-
Smirnoff statistic). However, some variables are better imputed
using alternative methods, e.g., Lung Sounds or Assistance
Devices.
Numerical variables were all unanimously imputed using the
MICE method. Moreover, the variables of medical importance
for cardiogenic shock classication (e.g., Heart Rate, O
2
ow,
Glucose, O
2
saturation) were all imputed using the MICE
method with relatively low Kolmogorov-Smirnoff statistic, and
p-values were in most cases not signicant, i.e., we could
conclude that MICE imputation provides us with imputed
variables that closely resemble original non-missing data.
Our results clearly show the superior ability of the MICE
method to reasonably impute data missing completely at
random, as in the case of the MIMIC III database. We also
suggest imputing more datasets, given the stochastic nature of
the imputation. Apart from assessing imputation quality,
multiple imputed datasets can be used in later stages for, e.g.,
sensitivity analysis, in which all datasets are used in grid search
for hyperparameter tuning or to increase the number of samples
for cross-validation of any diagnostic model. Although an
external cohort is critical to validate the model performance of
the medical model, cross-validation allows the estimation of the
prediction model error. It helps with optimizing the model and
classier selection. Obtaining external medical data for validation
is especially difcult due to the sensitive nature and associated
protection regulations, so thorough model testing and robust
results are usually prerequisites for establishing collaboration.
3.2. Cardiogenic shock prediction
We trained a classier on a subset of 9 clinically relevant
variables to test model performance on the imputed dataset. In
FIGURE 3
The number of signicant differences between original non-missing data and imputed data. Signicance was estimated using a 2-sample Kolmogorov-
Smirnoff test with p< 0.05. P-values are corrected for multiple comparisons using a Benjamini-Yekutieli FDR procedure (32). Shown are counts of
signicant differences per dataset (left panel) and variable (right panel).
Jajcay et al. 10.3389/fcvm.2023.1132680
Frontiers in Cardiovascular Medicine 06 frontiersin.org
the rst step, we performed a simple bivariate analysis with
appropriate statistical tests (chi-squared, unpaired t-test, or
Mann-Whitney U-test) for each variable. We chose the ones with
a proven or potential pathophysiological connection to
cardiogenic shock from the subset of variables with a statistically
signicant difference. In the next step, we narrowed down the
selection to only those variables that are available at the rst
contact with the patient:
Heart Rate
Blood glucose level
Oxygen saturation
O
2
ow of oxygen delivery device
Arterial blood pressure
Age
ECG classication of acute coronary syndrome
Sex
History of chronic heart failure
This cohort consisted of 2,253 patients (123 patients and 2,130
controls). The overview table of statistics in patient and control
cohorts is displayed in Table 3. We utilized gradient-boosted
trees for the classier type, representing a strong baseline for
these problems. In particular, we utilized a LightGBM (26)
implementation in python with traditional gradient-boosted
decision trees and 100 estimators, each using 31 leaves with
balanced class weight. Due to a relatively low incidence of
cardiogenic shock in patients with the ACS (between 5 and 10%)
(12), there was a relatively large class imbalance within our
cohort (approximately eight times more controls than patients).
To compensate for this fact (and after testing various methods
and techniques including manually setting class weights, or using
solely under- or over-sampling), we used a combination of over-
and under-sampling using the Synthetic Minority Over-sampling
Technique (SMOTE) algorithm (34) for over-sampling, followed
by Edited Nearest Neighbours (35) cleaning, as implemented in
the imblearn python package (36). The overall performance of
our trained model is summarized in Figure 5.
The average AUC for our trained model, as estimated using
repeated stratied K-Fold cross-validation technique with ve
splits and 50 repeats, reached 0.805 ± 0.039 (CI95% 0.7390.867).
The mean accuracy of our trained classier using the same cross-
validation technique reached 0.893 ± 0.014 (95% CI 0.8700.915).
FIGURE 4
Heatmap representing the match between imputed data and initially non-missing data for all variables and all imputed datasets. Variables are encodedas
rows, imputed datasets are encoded as columns, and color encodes the Kolmogorov-Smirnoff statistic as estimated using a 2-sample Kolmogorov-
Smirnoff test (lower is better), and stars mark signicance ( p< 0.05 *, p< 0.01 **, p< 0.001 ***). P-values are corrected for multiple comparisons
using a Benjamini-Yekutieli FDR procedure (33).
Jajcay et al. 10.3389/fcvm.2023.1132680
Frontiers in Cardiovascular Medicine 07 frontiersin.org
Considering all issues with missing data, class imbalance, and the
number of features used, this preliminary result is more than
acceptable and serves as a reasonable basis for further
improvement. After hyperparameter tuning or using a different
classier, we expect higher AUC and better performance.
4. Discussion
Our results provide a methodical pipeline for data pre-
processing for use in extensive EHR such as the MIMIC
database. Although some processing steps, such as patient
selection, are unique to this specic database, general data
processing strategies and imputation techniques are applicable in
most medical research working with large datasets.
We have used stochastic and non-stochastic imputation
methods in our pipeline to handle missing data. We relied on
Multiple Imputation by Chained Equations (MICE) (2325)as
our primary imputation algorithm. We included three additional
well-established imputation algorithms (KNN, Soft Impute, and
Iterative SVD) (37) to benchmark our stability and sensitivity
analysis. Multivariate techniques were chosen for their ability to
model interdependencies between variables, thus keeping the
covariance structure of the dataset. The evidence based on
extensive clinical and epidemiological trials is the cornerstone of
modern medicine. Although considerable efforts and delity are
put into preparation, data collection, and processing, but no
dataset is perfect, and missing and incomplete data are
unavoidable. Despite the potential of missing data to alter and
undermine the validity of research results, this problem has often
been overlooked in the medical literature (38). The study by
Wood et al. (39) demonstrated that the inadequate handling of
missing values with consequent impact on research results is a
common problem even in top-tier medical journals (including
BMJ, JAMA, Lancet, and New England Journal of Medicine.
Moreover, this study has shown that only 21% of the 71 trials
included a sensitivity analysis to inspect the quality of imputed
data.
Evaluating imputation quality is not well dened, and
universally accepted pipelines do not exist (31). In our work, we
opted for the visual assessment using heatmaps and violin plots
to estimate the characteristics of the imputation qualitatively and
for comparison of distributions of imputed data and original,
non-missing data by means of the Kolmogorov-Smirnov test (32)
with correction for multiple comparisons using a Benjamini-
Yekutieli FDR procedure (33) for distribution equivalent for the
quantitative assessment. In the pipeline we have studied, MICE
imputed datasets have shown a superior ability to impute
variables with multimodal distribution compared to other
methods. This methods stochastic nature allows imputing
multiple datasets and inspecting and comparing their variability.
Furthermore, testing the performance of diagnostic models
derived from multiple imputed datasets gives more robust results
thanks to hyperparameter tuning and increased samples for
cross-validation.
Another essential step in our imputation pipeline was the
expansion of our original dataset by including more patients and
variables. EHR include large quantities of data; usually, only a
subset of patients is selected for the specic research based on
the inclusion and exclusion criteria. Increasing sample size leads
to improved model performance. However, including all available
variables in the imputation model would signicantly increase
model complexity leading to a non-linear increase in
computational power needed (with a consequent increase in time
and resources needed) and may even lead to model overtting
(40). Selecting patients with similar proles and variables with
clinical and pathophysiological relationships to studied outcomes
may lead to optimization and improved model performance (22).
In our case, this methodology enabled us to create a model for
predicting CS in ACS patients, which would otherwise be
impossible due to the number and distribution of missing values.
The ability to identify high-risk patients prior to the
development of CS could allow to take pre-emptive measures,
such as the implantation of percutaneous mechanical circulatory
support, and thus prevent the development of CS leading to
improved survival. Predictive medicine is the future of
healthcare, ultimately leading to improved patient morbidity
mortality and cost reduction (41,42). Analysis of large EHR is
key in developing predictive medicine algorithms, so there is
denitely an emerging need for effective processing methodology.
We believe this proposed data processing pipeline offers good
instructions for analyzing sizeable electronic health records,
mainly focusing on managing missing data. Furthermore, it
TABLE 3 Summary of datasets.
no CS CS p-value
[n= 123] [n= 2130]
Sex (male) 64.46% 62.3% 0.698
Heart Rhythm <0.001
Sinus tachycardia 10.27% 28.69%
Sinus bradycardia 5.3% 1.64%
Ventricular tachycardia 0.14% 0.82%
1st degree AV block 0.85% 1.64%
Other 83.44% 67.21%
AH hist. (no) 47.47% 65.57% 0.001
chHF hist. (no) 64.32% 32.79% <0.001
Hypercholest. hist. (no) 72.93% 87.7% <0.001
EKG <0.001
Anterior STEMI or LBBB 17.61% 32.79%
Other STEMI 24.85% 31.15%
NSTEMI 47.18% 28.69%
Other 10.36% 7.38%
Heart Rate [bpm] 82.1 ± 16.343 94.1 ± 18.912 <0.001
Respiratory Rate [bpm] 17.1 ± 5.549 20.1 ± 5.917 <0.001
SaturationSpO2 [%] 97.8 ± 3.037 96.3 ± 4.699 0.001
Glucose [mg/dl] 157.8 ± 80.429 227.5 ± 135.491 <0.001
Systolic BP [mmHg] 125.6 ± 24.112 108.9 ± 20.538 <0.001
Age [y] 67.0 ± 12.401 71.0 ± 11.997 <0.001
Shock Index 0.69 ± 0.221 0.90 ± 0.285 <0.001
For continuous variables the table shows mean and standard deviation per dataset
and per group. For categorical variables, the table shows percentages of each. AH
hist., history of arterial hypertension, chHF hist., history of chronic heart failure,
Hyperchol. hist., history of hypercholesterolemia, EKG, electrocardiography, BP,
blood pressure, CS, cardiogenic shock.
Jajcay et al. 10.3389/fcvm.2023.1132680
Frontiers in Cardiovascular Medicine 08 frontiersin.org
offers good reproducibility and promotes further research using
different cohorts.
5. Limitations
Our pipeline was not tested on other datasets. Therefore, the
performance might differ in other EHR. Models were selected
based on available literature and team experience. Superior
computational power would allow imputing and analyzing more
datasets and include more models for analysis.
6. Conclusion
This proposed data processing pipeline offers good instructions
for analyzing sizeable EHR, mainly focusing on managing missing
data. Appropriate pre-processing with emphasis on handling of
missing data is crucial in analyzing large EHR.
Data availability statement
The raw data supporting the conclusions of this article will be
made available by the authors, without undue reservation.
Ethics statement
The studies involving human participants were reviewed
and approved by Institutional Review Board, The Chaim
Sheba Medical Center, Israel. Written informed consent
forparticipationwasnotrequiredforthisstudyin
FIGURE 5
Receiver operating characteristic curve for gradient boosted tree classier. Shown are all curves from repeated stratied K-Fold cross-validation using ve
splits and 50 repeats on all 10 MICE-imputed datasets (thin black lines) and mean ± standard deviation over all runs (thick red line). The classier scored
AUC 0.805 ± 0.039.
Jajcay et al. 10.3389/fcvm.2023.1132680
Frontiers in Cardiovascular Medicine 09 frontiersin.org
accordance with the national legislation and the institutional
requirements.
Author contributions
All authors listed have made a substantial, direct, and
intellectual contribution to the work. All authors contributed to
the article and approved the submitted version.
Funding
This research was partially supported by the Scientic Grant
Agency of the Ministry of Education, Science, Research and
Sport of the Slovak Republic grant (VEGA 1/0563/21).
Conict of interest
The authors declare that the research was conducted in the
absence of any commercial or nancial relationships that could
be construed as a potential conict of interest.
Publishers note
All claims expressed in this article are solely those of the
authors and do not necessarily represent those of their afliated
organizations, or those of the publisher, the editors and the
reviewers. Any product that may be evaluated in this article, or
claim that may be made by its manufacturer, is not guaranteed
or endorsed by the publisher.
References
1. Ghassemi M, Naumann T, Schulam P, Beam AL, Ranganath R. Opportunities in
machinelearning for healthcare.arXiv. (2018). arXiv:1806.00388. doi: arXiv:1806.00388v1
2. Bohm A, Jajcay N. Technical and practical aspects of articial intelligence in
cardiology. Bratisl Lek Listy. (2022) 123(0006-9248 (Print)):1621. doi: 10.4149/
BLL_2022_003
3. Nemethova A, Nemeth M, Michalconok G, Bohm A. Identication of kdd
problems from medical data. Adv Intell Syst Comput. Springer International
Publishing (2019) 985:1919. doi: 10.1007/978-3-030-19810-7_19
4. Sanchez-Martinez S, Camara O, Piella G, Cikes M, González-Ballester MÁ, Miron
M, et al. Machine learning for clinical decision-making: challenges and opportunities
in cardiovascular imaging. Front Cardiovasc Med. (2022) 8:765693. doi: 10.3389/fcvm.
2021.765693
5. Peterson E. Machine learning, predictive analytics, and clinical practice: can the
past inform the present? JAMA. (2019) 322(23):22834. doi: 10.1001/jama.2019.17831
6. Johnson AEW, Pollard TJ, Shen L, Lehman L-WH, Feng M, Ghassemi M, et al.
Mimic-Iii, a freely accessible critical care database. Sci Data. (2016) 3(1):160035.
doi: 10.1038/sdata.2016.35
7. Dai Z, Liu S, Wu J, Li M, Liu J, Li K. Analysis of adult disease characteristics and
mortality on mimic-iii. PLoS One. (2020) 15(4):e0232176. doi: 10.1371/journal.pone.
0232176
8. Song K, Guo C, Yang K, Li C, Ding N. Clinical characteristics of aortic aneurysm
in mimic-iii. Heart Surg Forum. (2021) 24(2):E351E8. doi: 10.1532/hsf.3571
9. Li F, Xin H, Zhang J, Fu M, Zhou J, Lian Z. Prediction model of in-hospital
mortality in intensive care unit patients with heart failure: machine learning-based,
retrospective analysis of the mimic-iii database. BMJ Open. (2021) 11(7):e044779.
doi: 10.1136/bmjopen-2020-044779
10. Peterkova A, Nemeth M, Bohm A. Computing missing values using neural
networks in medical eld.2018 IEEE 22nd international conference on intelligent
engineering systems (INES); 2018 21-23 June 2018
11. Bohm A, Jajcay N, Jankova J, Petrikova K, Bezak B. Articial intelligence model
for prediction of cardiogenic shock in patients with acute coronary syndrome. Eur
Heart J Acute Cardiovascular Care. (2022) 11(Supplement_1):i107. doi: 10.1093/
ehjacc/zuac041.077
12. De Luca L, Olivari Z, Farina A, Gonzini L, Lucci D, Di Chiara A, et al. Temporal
trends in the epidemiology, management, and outcome of patients with cardiogenic
shock complicating acute coronary syndromes. Eur J Heart Fail. (2015) 17
(11):112432. doi: 10.1002/ejhf.339
13. Mandawat A, Rao SV. Percutaneous mechanical circulatory support devices in
cardiogenic shock. Circ: Cardiovasc Interventions. (2017) 10(5):e004337. doi: 10.
1161/circinterventions.116.004337
14. Quan H, Sundararajan V, Halfon P, Fong A, Burnand B, Luthi J-C, et al. Coding
algorithms for dening comorbidities in Icd-9-Cm and Icd-10 administrative data.
Med Care. (2005) 43(11):11309. doi: 10.1097/01.mlr.0000182534.19832.83
15. Vincent J-L, Nielsen ND, Shapiro NI, Gerbasi ME, Grossman A, Doroff R, et al.
Mean arterial pressure and mortality in patients with distributive shock: a
retrospective analysis of the mimic-iii database. Ann Intensive Care. (2018) 8:107.
doi: 10.1186/s13613-018-0448-9
16. Lan P, Wang T-T, Li H-Y, Yan R-S, Liao W-C, Yu B-W, et al. Utilization of
echocardiography during septic shock was associated with a decreased 28-day
mortality: a propensity score-matched analysis of the mimic-iii database. Ann
Transl Med. (2019) 7(22):662. doi: 10.21037/atm.2019.10.79
17. Little RJA. A test of missing completely at random for multivariate data with
missing values. J Am Stat Assoc. (1988) 83(404):1198202. doi: 10.1080/01621459.
1988.10478722
18. Huque MH, Carlin JB, Simpson JA, Lee KJ. A comparison of multiple
imputation methods for missing data in longitudinal studies. BMC Med Res
Methodol. (2018) 18:168. doi: 10.1186/s12874-018-0615-6
19. Herbers J, Miller R, Walther A, Schindler L, Schmidt K, Gao W, et al. How to
deal with non-detectable and outlying values in biomarker research: best practices and
recommendations for univariate imputation approaches. Compr
Psychoneuroendocrinology. (2021) 7:100052. doi: 10.1016/j.cpnec.2021.100052
20. Waljee AK, Mukherjee A, Singal AG, Zhang Y, Warren J, Balis U, et al.
Comparison of imputation methods for missing laboratory data in medicine. BMJ
Open. (2013) 3(8):e002847. doi: 10.1136/bmjopen-2013-002847
21. He Y. Missing data analysis using multiple imputation: getting to the heart of the
matter. Circ Cardiovasc Qual Outcomes. (2010) 3(1):98105. doi: 10.1161/
circoutcomes.109.875658
22. Tang L, Song J, Belin TR, Unützer J. A comparison of imputation methods in a
longitudinal randomized clinical trial. Stat Med. (2006) 24(14):211128. doi: 10.1002/
sim.2099
23. Azur MJ, Stuart EA, Frangakis C, Leaf PJ. Multiple imputation by chained
equations: what is it and how does it work? Int J Methods Psychiatr Res. (2011) 20
(1):409. doi: 10.1002/mpr.329
24. Shah AD, Bartlett JW, Carpenter J, Nicholas O, Hemingway H. Comparison of
random forest and parametric imputation models for imputing missing data using
mice: a caliber study. Am J Epidemiol. (2014) 179(6):76474. doi: 10.1093/aje/kwt312
25. Ambler G, Omar RZ, Royston P. A comparison of imputation techniques for
handling missing predictor values in a risk model with a binary outcome. Stat
Methods Med Res. (2007) 16(3):27798. doi: 10.1177/0962280206074466
26. Ke G, Meng Q, Finley T, Wang T, Chen W, Ma W, et al. Lightgbm: A highly
efcient gradient boosting decision tree. Long Beach, CA, USA: NIPS (2017).
27. Morris TP, White IR, Royston P. Tuning multiple imputation by predictive
mean matching and local residual draws. BMC Med Res Methodol. (2014) 14(1):75.
doi: 10.1186/1471-2288-14-75
28. Malarvizhi RS, Thanamani AS. K-Nearest neighbor in missing data imputation.
IJERD. (2012) 5(1):57. doi: 10.9790/0661-0651215
29. Yao QK, James T. Accelerated and inexact soft-impute for large-scale matrix and
tensor completion. IEEE Trans Knowl Data Eng. (2018) 31(9):1. doi: 10.1109/tkde.
2018.2867533
30. Liu Y, Brown SD. Comparison of ve iterative imputation methods for
multivariate classication. Chemometr Intell Lab Syst. (2013) 120:10615. doi: 10.
1016/j.chemolab.2012.11.010
31. Salfrán D, Jordan P, Spiess M. Missing data: on criteria to evaluate imputation
methods. Hamburg: Universitat Hamburg (2016).
Jajcay et al. 10.3389/fcvm.2023.1132680
Frontiers in Cardiovascular Medicine 10 frontiersin.org
32. Abayomi K, Gelman A, Levy M. Diagnostics for multivariate imputations. J
R Stat Soc, C: Appl Stat. (2008) 57(3):27391. doi: 10.1111/j.1467-9876.2007.00613.x
33. Stevens JR, Al Masud A, Suyundikov A. A comparison of multiple testing
adjustment methods with block-correlation positively-dependent tests. PLoS One.
(2017) 12(4):e0176124. doi: 10.1371/journal.pone.0176124
34. Pears RF, Finlay J, Connor AM. Synthetic minority over-sampling technique
(smote) for predicting software build outcomes. arXiv. 1407.2330 (2014). doi: 10.
48550/arxiv.1407.2330
35. Alejo R, Sotoca JM, Valdovinos RM, Toribio P. Edited nearest neighbor rule for
improving neural networks classications. In: Zhang L, Lu B-L, Kwok J, editors.
Advances in neural networks.ISNN 2010; 2010 2010//. Berlin, Heidelberg: Springer
Berlin Heidelberg. (2010). p. 30310.
36. Kovács G. Smote-Variants: a python implementation of 85 minority oversampling
techniques. Neurocomputing. (2019) 366:3524. doi: 10.1016/j.neucom.2019.06.100
37. Rafsunjani SS, Safa RS, Imran A, Rahim S, Nandi D. An empirical comparison of
missing value imputation techniques on aps failure prediction. IJ Inf Technol Comput
Sci. (2019) 11:219. doi: 10.5815/ijitcs.2019.02.03
38. Sterne JAC, White IR, Carlin JB, Spratt M, Royston P, Kenward MG, et al.
Multiple imputation for missing data in epidemiological and clinical research:
potential and pitfalls. Br Med J. (2009) 338:b2393. doi: 10.1136/bmj.b2393
39. Wood AM, White IR, Thompson SG. Are missing outcome data adequately
handled? A review of published randomized controlled trials in Major medical
journals. Clin Trials. (2004) 1(4):36876. doi: 10.1191/1740774504cn032oa
40. Noghrehchi F, Stoklosa J, Penev S, Warton DI. Selecting the model for multiple
imputation of missing data: just use an ic!. Stat Med. (2021) 40(10):246797. doi: 10.
1002/sim.8915
41. Staartjes VE, de Wispelaere MP, Vandertop WP, Schröder ML. Deep learning-
based preoperative predictive analytics for patient-reported outcomes following
lumbar discectomy: feasibility of center-specic modeling. Spine J. (2019) 19
(5):85361. doi: 10.1016/j.spinee.2018.11.009
42. Alonso SG, de la Torre Díez I, Zapiraín BG. Predictive, personalized,
preventive and participatory (4p) medicine applied to telemedicine and
ehealth in the literature. J Med Syst. (2019) 43(5):140. doi: 10.1007/s10916-019-
1279-4
Jajcay et al. 10.3389/fcvm.2023.1132680
Frontiers in Cardiovascular Medicine 11 frontiersin.org
... Our systematic review included four retrospective studies [16][17][18][19] and two prospective cohort studies [20,21]. All studies were conducted in the USA except for Bai et al. 2021 [21], which was conducted in China. ...
... All studies were conducted in the USA except for Bai et al. 2021 [21], which was conducted in China. Rahman et al. 2022 [19] included patients hospitalized with acute decompensated heart failure, whereas two studies [16,18] included patients having cardiac catheterizations for acute coronary syndrome. The remaining three studies [17,20,21] included patients from various backgrounds, including those with heart failure and myocardial infarction. ...
... Area under the curve (AUC): Three studies reported [16,18,20] the AUC in a way to be combined in a metaanalysis with a total sample size of 10,869. The pooled mean of AUC was 0.808 (95% CI: 0.727, 0.890) (Figure 2). ...
Article
Full-text available
Cardiogenic shock (CS) may have a negative impact on mortality in patients with heart failure (HF) or acute myocardial infarction (AMI). Early prediction of CS can result in improved survival. Artificial intelligence (AI) through machine learning (ML) models have shown promise in predictive medicine. Here, we conduct a systematic review and meta-analysis to assess the effectiveness of these models in the early prediction of CS. A thorough search of the PubMed, Web of Science, Cochrane, and Scopus databases was conducted from the time of inception until November 2, 2023, to find relevant studies. Our outcomes were area under the curve (AUC), the sensitivity and specificity of the ML model, the accuracy of the ML model, and the predictor variables that had the most impact in predicting CS. Comprehensive Meta-Analysis (CMA) Version 3.0 was used to conduct the meta-analysis. Six studies were considered in our study. The pooled mean AUC was 0.808 (95% confidence interval: 0.727, 0.890). The AUC in the included studies ranged from 0.77 to 0.91. ML models performed well, with accuracy ranging from 0.88 to 0.93 and sensitivity and specificity of 58%-78% and 88%-93%, respectively. Age, blood pressure, heart rate, oxygen saturation, and blood glucose were the most significant variables required by ML models to acquire their outputs. In conclusion, AI has the potential for early prediction of CS, which may lead to a decrease in the high mortality rate associated with it. Future studies are needed to confirm the results.
... In several recent studies, gradient boosting-based or logistic regression models have been used to predict the occurrence of all-cause CS using retrospective data from different US datasets based on ICD codes. [42][43][44] Another study trained a lasso model to predict late-onset CS after ST-elevation myocardial infarction. 45 Although the accuracy of the ML models in all of these reports is promising, none of them have provided an adequate external validation cohort; thus, their external validity remains to be shown prior to any considerations of their use in clinical practice. ...
... In addition to pharmacological therapies and biomarkers, the integration of artificial intelligence (AI) into HF diagnostic models holds significant promise for the future of HF management. AI-based algorithms, particularly machine learning and deep learning techniques are being leveraged to analyze complex datasets, including electronic health records, medical imaging, and genetic information, to identify patterns and predict heart failure outcomes [89]. These AI-driven diagnostic models have the potential to enhance early detection and risk stratification, ultimately leading to improved clinical outcomes and resource allocation within healthcare systems. ...
Preprint
Full-text available
Heart failure (HF) is a growing issue in developed countries and most patients present with a reduced ejection fraction. This condition significantly impairs quality of life and represents a threat not only for patients, but also for the entire healthcare system due to its high management cost. In the last decade, new medical treatments and devices have been developed to reduce HF hospitalizations and improve prognosis, while reducing the overall mortality rate. Medical therapy does not achieve the same results in all patients and, whenever it fails, technology comes in handy: novel devices for the management of HF have reduced symptoms. They have also improved the treatment of fluid retention and life threatening arrhythmias. The present review article gives to the readers a comprehensive overview of the most recent and important findings that need to be considered in HF with reduced ejection fraction patients. Both novel medical treatments and devices have been presented and discussed.
... In addition to pharmacological therapies and biomarkers, the integration of artificial intelligence (AI) into HF diagnostic models holds significant promise for the future of HF management. AI-based algorithms, particularly machine learning and deep learning techniques, are being leveraged to analyze complex datasets, including electronic health records, medical imaging and genetic information, to identify patterns and predict heart failure outcomes [98]. These AI-driven diagnostic models have the potential to enhance early detection and risk stratification, ultimately leading to improved clinical outcomes and resource allocation within healthcare systems. ...
Article
Full-text available
Heart failure (HF) is a growing issue in developed countries; it is often the result of underlying processes such as ischemia, hypertension, infiltrative diseases or even genetic abnormalities. The great majority of the affected patients present a reduced ejection fraction (≤40%), thereby falling under the name of “heart failure with reduced ejection fraction” (HFrEF). This condition represents a major threat for patients: it significantly affects life quality and carries an enormous burden on the whole healthcare system due to its high management costs. In the last decade, new medical treatments and devices have been developed in order to reduce HF hospitalizations and improve prognosis while reducing the overall mortality rate. Pharmacological therapy has significantly changed our perspective of this disease thanks to its ability of restoring ventricular function and reducing symptom severity, even in some dramatic contexts with an extensively diseased myocardium. Notably, medical therapy can sometimes be ineffective, and a tailored integration with device technologies is of pivotal importance. Not by chance, in recent years, cardiac implantable devices witnessed a significant improvement, thereby providing an irreplaceable resource for the management of HF. Some devices have the ability of assessing (CardioMEMS) or treating (ultrafiltration) fluid retention, while others recognize and treat life-threatening arrhythmias, even for a limited time frame (wearable cardioverter defibrillator). The present review article gives a comprehensive overview of the most recent and important findings that need to be considered in patients affected by HFrEF. Both novel medical treatments and devices are presented and discussed.
Article
Aims Cardiogenic shock (CS) is a severe complication of acute coronary syndrome (ACS) with mortality rates approaching 50%. The ability to identify high-risk patients prior to the development of CS may allow for pre-emptive measures to prevent the development of CS. The objective was to derive and externally validate a simple, machine learning (ML)-based scoring system using variables readily available at first medical contact to predict the risk of developing CS during hospitalization in patients with ACS. Methods and results Observational multicentre study on ACS patients hospitalized at intensive care units. Derivation cohort included over 40 000 patients from Beth Israel Deaconess Medical Center, Boston, USA. Validation cohort included 5123 patients from the Sheba Medical Center, Ramat Gan, Israel. The final derivation cohort consisted of 3228 and the final validation cohort of 4904 ACS patients without CS at hospital admission. Development of CS was adjudicated manually based on the patients’ reports. From nine ML models based on 13 variables (heart rate, respiratory rate, oxygen saturation, blood glucose level, systolic blood pressure, age, sex, shock index, heart rhythm, type of ACS, history of hypertension, congestive heart failure, and hypercholesterolaemia), logistic regression with elastic net regularization had the highest externally validated predictive performance (c-statistics: 0.844, 95% CI, 0.841–0.847). Conclusion STOP SHOCK score is a simple ML-based tool available at first medical contact showing high performance for prediction of developing CS during hospitalization in ACS patients. The web application is available at https://stopshock.org/#calculator.
Article
Full-text available
Background: Artificial intelligence (AI) can radically change almost every aspect of the human experience. In the medical field, there are numerous applications of AI and subsequently, in a relatively short time, significant progress has been made. Cardiology is not immune to this trend, this fact being supported by the exponential increase in the number of publications in which the algorithms play an important role in data analysis, pattern discovery, identification of anomalies, and therapeutic decision making. Furthermore, with technological development, there have appeared new models of machine learning (ML) and deep learning (DP) that are capable of exploring various applications of AI in cardiology, including areas such as prevention, cardiovascular imaging, electrophysiology, interventional cardiology, and many others. In this sense, the present article aims to provide a general vision of the current state of AI use in cardiology. Results: We identified and included a subset of 200 papers directly relevant to the current research covering a wide range of applications. Thus, this paper presents AI applications in cardiovascular imaging, arithmology, clinical or emergency cardiology, cardiovascular prevention, and interventional procedures in a summarized manner. Recent studies from the highly scientific literature demonstrate the feasibility and advantages of using AI in different branches of cardiology. Conclusions: The integration of AI in cardiology offers promising perspectives for increasing accuracy by decreasing the error rate and increasing efficiency in cardiovascular practice. From predicting the risk of sudden death or the ability to respond to cardiac resynchronization therapy to the diagnosis of pulmonary embolism or the early detection of valvular diseases, AI algorithms have shown their potential to mitigate human error and provide feasible solutions. At the same time, limits imposed by the small samples studied are highlighted alongside the challenges presented by ethical implementation; these relate to legal implications regarding responsibility and decision making processes, ensuring patient confidentiality and data security. All these constitute future research directions that will allow the integration of AI in the progress of cardiology.
Article
Aims Myocardial infarction and heart failure are major cardiovascular diseases that affect millions of people in the USA with morbidity and mortality being highest among patients who develop cardiogenic shock. Early recognition of cardiogenic shock allows prompt implementation of treatment measures. Our objective is to develop a new dynamic risk score, called CShock, to improve early detection of cardiogenic shock in the cardiac intensive care unit (ICU). Methods and results We developed and externally validated a deep learning-based risk stratification tool, called CShock, for patients admitted into the cardiac ICU with acute decompensated heart failure and/or myocardial infarction to predict the onset of cardiogenic shock. We prepared a cardiac ICU dataset using the Medical Information Mart for Intensive Care-III database by annotating with physician-adjudicated outcomes. This dataset which consisted of 1500 patients with 204 having cardiogenic/mixed shock was then used to train CShock. The features used to train the model for CShock included patient demographics, cardiac ICU admission diagnoses, routinely measured laboratory values and vital signs, and relevant features manually extracted from echocardiogram and left heart catheterization reports. We externally validated the risk model on the New York University (NYU) Langone Health cardiac ICU database which was also annotated with physician-adjudicated outcomes. The external validation cohort consisted of 131 patients with 25 patients experiencing cardiogenic/mixed shock. CShock achieved an area under the receiver operator characteristic curve (AUROC) of 0.821 (95% CI 0.792–0.850). CShock was externally validated in the more contemporary NYU cohort and achieved an AUROC of 0.800 (95% CI 0.717–0.884), demonstrating its generalizability in other cardiac ICUs. Having an elevated heart rate is most predictive of cardiogenic shock development based on Shapley values. The other top 10 predictors are having an admission diagnosis of myocardial infarction with ST-segment elevation, having an admission diagnosis of acute decompensated heart failure, Braden Scale, Glasgow Coma Scale, blood urea nitrogen, systolic blood pressure, serum chloride, serum sodium, and arterial blood pH. Conclusion The novel CShock score has the potential to provide automated detection and early warning for cardiogenic shock and improve the outcomes for millions of patients who suffer from myocardial infarction and heart failure.
Article
Full-text available
Funding Acknowledgements Type of funding sources: Public grant(s) – National budget only. Main funding source(s): Ministry of Education, Science, Research and Sport of the Slovak Republic VEGA Background Cardiogenic shock (CS) is a serious life-threatening condition affecting almost 10% of patients suffering from acute coronary syndrome (ACS). Despite recent treatment improvements such as mechanical circulatory support devices (MCS), mortality remains 50%. It is hypothesized that early implantation of MCS before hemodynamic deterioration could prevent CS. Therefore, knowing which ACS patient will progress into CS would be of paramount importance. Purpose We aimed to develop a model based on machine learning algorithms for CS prediction in patients with ACS. Method Over 40 000 patients from critical care units of the Beth Israel Deaconess Medical Center database were extensively analyzed and sorted. Patients suffering from acute coronary syndrome undergoing cardiac catheterizations were selected and divided into two groups based on the development of CS. Patients in CS at the time of admission were excluded. The study population consisted of 3056 patients who didn’t develop CS and of 176 patients who did develop CS. Important information was also extracted manually from textual summaries of hospital stays. Potentially relevant clinical variables for shock prediction were selected using supervised feature selection, and missing values were supplemented using imputation methods. Seven well-known and established machine learning algorithms were used. Based on preliminary evaluation of classifier performance (AUC on random train-test split with 30% test data), we selected two best-performing algorithms: Logistic Regression and Gaussian Process classifier with Radial Basis Function (RBF) kernel. Both models were subsequently validated using Repeated Stratified K-Fold cross-validation with 5 folds and 20 repeats each. Results Age, heart rate, mean arterial pressure, respiratory rate, oxygen flow (liters per minute), peripheral oxygen saturation, blood glucose, pain type, heart rhythm and ectopy frequency were chosen as input variables. Both models showed good discrimination. The Logistic Regression model scored AUC of 0.76 ± 0.04. The Gaussian Process classifier with Radial Basis Function (RBF) kernel scored AUC of 0.77 ± 0.03. Conclusion According to our knowledge this is the first study that uses machine learning algorithms to predict CS in patients with ACS based on easily obtainable clinical variables. Further explorations and refinement of used imputation, feature selection, machine learning techniques and validation on an external cohort might result in even better performance of the proposed prediction models. Furthermore, these prediction models could be transformed into a simple predictive scoring system available in clinical practice.
Article
Full-text available
The use of machine learning (ML) approaches to target clinical problems is called to revolutionize clinical decision-making in cardiology. The success of these tools is dependent on the understanding of the intrinsic processes being used during the conventional pathway by which clinicians make decisions. In a parallelism with this pathway, ML can have an impact at four levels: for data acquisition, predominantly by extracting standardized, high-quality information with the smallest possible learning curve; for feature extraction, by discharging healthcare practitioners from performing tedious measurements on raw data; for interpretation, by digesting complex, heterogeneous data in order to augment the understanding of the patient status; and for decision support, by leveraging the previous steps to predict clinical outcomes, response to treatment or to recommend a specific intervention. This paper discusses the state-of-the-art, as well as the current clinical status and challenges associated with the two later tasks of interpretation and decision support, together with the challenges related to the learning process, the auditability/traceability, the system infrastructure and the integration within clinical processes in cardiovascular imaging.
Article
Full-text available
Objective The predictors of in-hospital mortality for intensive care units (ICUs)-admitted heart failure (HF) patients remain poorly characterised. We aimed to develop and validate a prediction model for all-cause in-hospital mortality among ICU-admitted HF patients. Design A retrospective cohort study. Setting and participants Data were extracted from the Medical Information Mart for Intensive Care (MIMIC-III) database. Data on 1177 heart failure patients were analysed. Methods Patients meeting the inclusion criteria were identified from the MIMIC-III database and randomly divided into derivation (n=825, 70%) and a validation (n=352, 30%) group. Independent risk factors for in-hospital mortality were screened using the extreme gradient boosting (XGBoost) and the least absolute shrinkage and selection operator (LASSO) regression models in the derivation sample. Multivariate logistic regression analysis was used to build prediction models in derivation group, and then validated in validation cohort. Discrimination, calibration and clinical usefulness of the predicting model were assessed using the C-index, calibration plot and decision curve analysis. After pairwise comparison, the best performing model was chosen to build a nomogram according to the regression coefficients. Results Among the 1177 admissions, in-hospital mortality was 13.52%. In both groups, the XGBoost, LASSO regression and Get With the Guidelines-Heart Failure (GWTG-HF) risk score models showed acceptable discrimination. The XGBoost and LASSO regression models also showed good calibration. In pairwise comparison, the prediction effectiveness was higher with the XGBoost and LASSO regression models than with the GWTG-HF risk score model (p<0.05). The XGBoost model was chosen as our final model for its more concise and wider net benefit threshold probability range and was presented as the nomogram. Conclusions Our nomogram enabled good prediction of in-hospital mortality in ICU-admitted HF patients, which may help clinical decision-making for such patients.
Article
Full-text available
Non-detectable (ND) and outlying concentration values (OV) are a common challenge of biomarker investigations. However, best practices on how to aptly deal with the affected cases are still missing. The high methodological heterogeneity in biomarker-oriented research, as for example, in the field of psychoneuroendocrinology, and the statistical bias in some of the applied methods may compromise the robustness, comparability, and generalizability of research findings. In this paper, we describe the occurrence of ND and OV in terms of a model that considers them as censored data, for instance due to measurement error cutoffs. We then present common univariate approaches in handling ND and OV by highlighting their respective strengths and drawbacks. In a simulation study with lognormal distributed data, we compare the performance of six selected methods, ranging from simple and commonly used to more sophisticated imputation procedures, in four scenarios with varying patterns of censored values as well as for a broad range of cutoffs. Especially deletion, but also fixed-value imputations bear a high risk of biased and pseudo-precise parameter estimates. We also introduce censored regressions as a more sophisticated option for a direct modeling of the censored data. Our analyses demonstrate the impact of ND and OV handling methods on the results of biomarker-oriented research, supporting the need for transparent reporting and the implementation of best practices. In our simulations, the use of imputed data from the censored intervals of a fitted lognormal distribution shows preferable properties regarding our established criteria. We provide the algorithm for this favored routine for a direct application in R on the Open Science Framework (https://osf.io/spgtv). Further research is needed to evaluate the performance of the algorithm in various contexts, for example when the underlying assumptions do not hold. We conclude with recommendations and potential further improvements for the field.
Article
Full-text available
Multiple imputation and maximum likelihood estimation (via the expectation- maximization algorithm) are two well-known methods readily used for analyzing data with missing values. While these two methods are often considered as being distinct from one another, multiple imputation (when using improper imputation) is actually equivalent to a stochastic expectation-maximization approximation to the likelihood. In this paper we exploit this key result to show that familiar likelihood-based approaches to model selection, such as Akaike's Information Criterion (AIC) and the Bayesian Information Criterion (BIC), can be used to choose the imputation model that best fits the observed data. Poor choice of imputation model is known to bias inference, and while sensitivity analysis has often been used to explore the implications of different imputation models, we show that the data can be used to choose an appropriate imputation model via conventional model selection tools. We show that BIC can be consistent for selecting the correct imputation model in the presence of missing data. We verify these results empirically through simulation studies, and demonstrate their practicality on two classical missing data examples. An interesting result we saw in simulations was that not only can parameter estimates be biased by misspecifying the imputation model, but also by overfitting the imputation model. This emphasises the importance of using model selection not just to choose the appropriate type of imputation model, but also to decide on the appropriate level of imputation model complexity.
Article
Full-text available
Purpose To deeply analyze the basic information and disease information of adult patients in the MIMIC-III (Medical Information Mart for Intensive Care III) database, and provide data reference for clinicians and researchers. Materials and methods Tableau2019.1.0 and Navicat12.0.29 were used for data analysis and extraction of disease distribution of adult patients in the MIMIC-III database. Result A total of 38,163 adult patients were included in the MIMIC-III database. Only 38,156 patients with the first diagnosis were selected. Among them, 21,598 were males accounting for 56.6% the median age was 66 years (Q1-Q3: 53–78), the median length of a hospital stay was 7 days (Q1-Q3: 4–12), and the median length of an ICU stay was 2.1 days (Q1-Q3: 1.2–4.1). Septicemia was the disease with the highest mortality rate among patients and the total mortality rate was 48.9%. The disease with the largest number of patients at the last time was other forms of chronic ischemic heart disease. Conclusion By analyzing the patients’ basic information, the admission spectrum and the disease morbidity and mortality can help more researchers understand the MIMIC-III database and facilitate further research.
Article
Artificial intelligence (AI) is here to stay. It is not a future anymore, and there are many particular problems in cardiology that are already being solved via machine learning (ML), and many more are to come. AI cannot solve complex tasks yet, and probably this will not change in the upcoming years. Therefore, cardiologists do not have to be afraid that computers will replace them. However, cardiologists who will not be able to use ML algorithms in their clinical practice will be replaced by those who will. (Fig. 2, Ref. 50). Keywords: artificial intelligence, cardiology, potential machine learning, survival models, classification algorithms, computer vision, automated analysis of various imaging examinations, ECG interpretation, phenotype clustering, pathophysiological mechanisms.
Article
The use of machine learning (ML) in healthcare raises numerous ethical concerns, especially as models can amplify existing health inequities. Here, we outline ethical considerations for equitable ML in the advancement of healthcare. Specifically, we frame ethics of ML in healthcare through the lens of social justice. We describe ongoing efforts and outline challenges in a proposed pipeline of ethical ML in health, ranging from problem selection to postdeployment considerations. We close by summarizing recommendations to address these challenges.
Article
Background: Aortic aneurysm (AA) is an aortic disorder prone to sudden, life-threatening aortic dissection or rupture, with poor clinical outcomes. In this study, we aimed to analyze the clinical characteristics of AA in MIMIC-III to explore implications for management. Methods: All patients with AA, including abdominal aortic aneurysm (AAA) and thoracic aortic aneurysm (TAA), in the MIMIC-III database were included. Clinical and laboratory variables were analyzed and compared in AAA and TAA. Results: A total of 345 patients, including 183 patients with AAA and 162 patients with TAA, were enrolled in this study. The in-hospital mortality in AAA and TAA groups was 6.01% and 3.7%, respectively. In the nonsurvivor groups in both AAA and TAA, patients were older, and the incidence of surgery was lower. In the nonsurvivor group of AAA, the levels of alanine aminotransferase, aspartate aminotransferase, urea nitrogen, creatinine, lactate dehydrogenase (LDH), creatine kinase, anion gap, and lactate were significantly higher in the nonsurvivor group, whereas the level of albumin was lower. In the nonsurvivor group of TAA, the level of LDH significantly increased and the level of albumin decreased. In the nonsurgery group, in-hospital mortality was higher, and patients were older, with higher levels of glucose, total bilirubin, urea nitrogen, and creatinine and longer length of stay in ICU and hospital. Conclusion: Age, surgery, albumin, and LDH showed significant differences between survivor and nonsurvivor groups in both AAA and TAA. In the nonsurgery group, the mean age was older and disease severity was worse, with poorer clinical outcomes. Older AA patients without surgery and with lower levels of albumin and higher levels of LDH had higher risk of in-hospital mortality.
Article
Modern electronic health records (EHRs) provide data to answer clinically meaningful questions. The growing data in EHRs makes healthcare ripe for the use of machine learning. However, learning in a clinical setting presents unique challenges that complicate the use of common machine learning methodologies. For example, diseases in EHRs are poorly labeled, conditions can encompass multiple underlying endotypes, and healthy individuals are underrepresented. This article serves as a primer to illuminate these challenges and highlights opportunities for members of the machine learning community to contribute to healthcare.