Conference PaperPDF Available

Learning Insurance Benefit Rules from Policy Texts with Small Labeled Data

Authors:

Abstract and Figures

To protect vital health program funds from being paid out on services that are wasteful and inconsistent with medical practices, government healthcare insurance programs need to validate the integrity of claims submitted by providers for reimbursement. However, due the complexity of healthcare billing policies and the lack of coded rules, maintaining "integrity" is a labor-intensive task, often narrow-scope and expensive. We propose an approach that combines deep learning and an ontology to support the extraction of actionable knowledge on benefit rules from regulatory healthcare policy text. We demonstrate its feasibility even in the presence of small ground truth labeled data provided by policy investigators. Leveraging deep learning and rich ontological information enables the system to learn from human corrections and capture better benefit rules from policy text, beyond just using a deterministic approach based on pre-defined textual and semantic pattterns.
Content may be subject to copyright.
Learning Insurance Benefit Rules from Policy Texts with Small Labeled Data
Gabriele Picco, Hoang Thanh Lam, Vanessa Lopez, John Segrave-Daly, Miao Wei, Marco Sbodio, Inge Vejsbjerg,
Seamus Brady and Morten Kristiansen
IBM Research Europe and IBM Watson Health, Dublin, Ireland
Abstract
To protect vital health program funds from being paid out
on services that are wasteful and inconsistent with
medical practices, government healthcare insurance
programs need to validate the integrity of claims
submitted by providers for reimbursement. However, due
the complexity of healthcare billing policies and the lack
of coded rules, maintaining integrity is a labor-
intensive task, often narrow-scope and expensive. We
propose an approach that combines deep learning and an
ontology to support the extraction of actionable
knowledge on benefit rules from regulatory healthcare
policy text. We demonstrate its feasibility even in the
presence of small ground truth labeled data provided by
policy investigators. Leveraging deep learning and rich
ontological information enables the system to learn from
human corrections and capture better benefit rules from
policy text, beyond just using a deterministic approach
based on pre-defined textual and semantic pattterns.
Keywords: health policy, deep learning, ontology
Introduction
Health and Social care Programs have a major impact on
outcomes for vulnerable citizens. To ensure resources are
distributed fairly to citizens, and are not lost to Fraud,
Waste or Abuse (FWA), these Programs typically have
large bodies of policy text describing benefit rules. In
practice, however, only a fraction of these rules get
automated and the lack of coded rules creates major
problems down-stream. Firstly, regulators tasked with
protecting the Programs financial integrity cannot get an
overview of the compliance landscapeto identify FWA
and prioritize investigations. Secondly, FWA detection is
typically based on statistical and data-analytic approaches
[1][2], making difficult for investigators to determine
which (if any) policy may have been violated. Finally,
those being regulated (Providers) have no fast/automatic
way to check the compliance of their claims, leading to
friction, cost and avoidable errors.
Recently, the Rules as Codemovement [3] has emerged
as a response to these issues. It calls for policy rules to be
published as digital twins”— in both human and
machine-consumable form making them amenable to
automated compliance-checking, fairness-checking,
1
Ontology and benefit rules to be made available in
https://github.com/IBM/rules_extraction_from_healthcare_policy
loophole identification, and what-if analysis, as well as
building trust through transparency.
In this vein, our prior work [4][5] describes a
deterministic approach to extracting “benefit rules” from
healthcare policy, in a form that is both human and
machine consumable. Each rule is associated to the policy
text from where it was extracted, which facilitates easy
review/validation and correction by medically-aware
users before use.
In this paper, we move this approach forward by
proposing a model to (1) learn from the
addition/correction of new rules and (2) identify new
elements, beyond the vocabularyon which the model
was trained.
To achieve this, we combine neural and symbolic
approaches. A neural NLP model learns which spans, i.e.,
one or more consecutive words, in policy text are rule
conditions— e.g. services that are mutually exclusive
(not payable when billed together). A domain ontology
built with investigators [5], supports both human
understanding/oversight and acts as a blueprint for
constructing semantically-meaningful rules from the
conditions labeled by the neural model e.g.,
determining if a complete mutually exclusiverule can
be formed. To train the NLP model, we use a small set of
141 benefit rules identified by professional investigators
in dental policy documents from two US states
1
. For each
rule, we annotate the rule text with ontology-aligned
labels describing the different condition types. Two
models are then trained: first, a classifier model to assess
which paragraphs contain benefit rules (many do not), and
second a model to predict and label specific spans.
We evaluate results against a gold standard of policy rules
obtained from professional investigators and compare the
results against the existing deterministic baseline,
presented in [5] and find that this fusion of neural and
symbolic approaches (1) improves extraction
performance, with a small up-front labeling cost and a
surprisingly small set of benefit rules provided by
investigators; and (2) correctly labels previously-unseen
elements in rule conditions e.g. healthcare services that
are not (yet) present in our ontology or terminology. In
short, this approach shows good promise for lowering
extraction cost and improving rule quality, ultimately
helping investigators to review health and social care
policies and support the extraction of actionable rules.
Methods
Benefit Rule Extraction Pipeline
In this section, we briefly describe the existent pipeline to
extract benefit rules from healthcare policy. First, an off-
the-shelf PDF conversion tool [6] is used to obtain an
HTML representation with headings, paragraphs and
sentences. These paragraphs are filtered via a BERT-
based classifier. No other text pre-processing or cleaning
is performed. The classifier simply filters out policy
paragraphs that do not appear to contain any rules, thereby
reducing noise, time, and computational load for the rest
of the pipeline.
Each paragraph may contain zero, one, or more benefit
rules. An example of a paragraph with two different
benefit rules types can be seen in Figure 1. While the
wording of policy texts can differ significantly between
geographical regions or policy topics, all policies set out
similar guidelines based on common compliance
concepts such as eligible patients (e.g., based on age,
medical history), billable places of service (e.g., home,
hospital), maximum billable units of service or monetary
amounts per member in a given period, services that
should not be billed together, etc. This shared
conceptualization behind benefit rules is captured in a
domain ontology. The ontology is used to ensure
semantically-meaningful rules are extracted from the
policy text. For example, the property applicable
serviceexpects a type of billed service as its value, or
that there can only be one applicable time period per
benefit rule (max cardinality of 1). The ontology also
links to other data sources with relevant terminology,
such as codes for billable dental services [8] or places of
service [9], and guides the annotation and extraction of
relevant domain entities, relationships and the logical
constraints that underpin them.
Expert-labeled datasets are expensive to develop (and
necessarily small), therefore a deterministic approach to
extract rules from policy texts was proposed in [5], where
after annotating the mentions of entities and relations
based on the ontology, NLP tools are applied to identify
functional dependencies between the annotated
ontological terms as well as their semantic role in the
sentence (actions, agent, theme of the action, polarity,
etc.). Then, based on a combination of linguistic rules and
domain-independent semantic patterns to reason over the
ontology, textual dependencies are translated into
meaningful benefit rules. For example, a number will be
interpreted as a unit, amount of time, or age limitation by
considering its semantic proximity to an annotated
property for an age range, a unit limit, or time period (both
expecting a number as the range).
The extracted benefit rules are presented to the user in the
form of editable condition entity/value pairs, that enable
investigators to review the extracted rule against the
policy text and correct extraction errors and omissions.
These validated benefit rules form a shared store of high-
quality, machine and human readable rules, making for a
more transparent rule creation and consumption.
However, the deterministic extraction is unable to learn
from those human corrections, and the approach is limited
by, first, the coverage of the ontology annotators and the
need to map policy text to known ontology entities; and
second, the use of a rule-based approach to identify
linguistic dependencies and semantic patterns among
those relevant entities that can be transformed into
semantically meaningful connections to build a benefit
rule.
Figure 1: Benefit rules extracted from a paragraph in a dental policy
[7]. On top, a Mutually Exclusive rule on services that cannot be billed
together in a given period. On the bottom, a Service Limitation rule on
the units of service a provider may bill per member over a time period
We propose a deep learning aproach to predict textual
spans that constitute conditions of a benefit rule. That is,
not just saying that a comprehensive oral evaluationis
an instance of a service, but that in the context of the
sentence it has two distinct roles simultaneously: an
applicable service and a mutually exclusive non-
reimbursable service, for a Service Limitation rule and
Mutually Exclusive rule respectively.
Benefit Rule Span Prediction
We consider the rule extraction as a span prediction
problem. Let
𝒔 = 𝒕𝟏𝒕𝟐… 𝒕𝒎
be a sentence with m tokens
and V the set of labels. Each token t is associated with one
or more labels from V. Assume that we have k labels in
our problem, i.e. |V| = k, e.g., for our policy rule
extraction, we have k=27 labels with examples provided
in Figure 2 which shows sentences with associated labels.
A span can have more than one label. Therefore we use
BERT [10] as a backbone network and we add a multi-
class classification head on the top of it to enable multi-
class classification. We trained the models with Adam
optimization [13], mini-batch size 8, learning rate
0.00001 and 100 epochs.
Figure 2 Data augmentation: given an input labeled sentence, a new
sentence is created by randomly sampling the spans from the set of
spans with the same labels in V.
Data Augmentation. Investigators created benefit rules
that were used to manually label span fragments in 141
policy paragraphs using the 27 labels in V, one for each
benefit rule condition defined in our ontology, that our
deep learning models can work on. The label schema was
selected to fully cover all relevant conditions that may be
included in a benefit rule, but also that is simple enough
for human annotators. While this process can be
automated to some extent, human supervision is needed
to match the values in a benefit rule to spans in the
sentence. Our main challenge during the labeling process
is the difference in the way annotators interpret the
concepts in an unfamiliar domain. Time-consuming
cross-checks were carried out to ensure the consistency
between labels provided by different annotators in the
team. We ensured each label was reviewed by at least two
annotators based on a set of annotation gudeliness to
ensure consistency. For example, the label billing-
commonality covers the span per dental practice
(including the preposition per”), as billing-commonality
is the condition used by domain experts to indicate that
the claims need to be billed by the same provider.
Given the fact that acquisition of labeled data for a new
policy is expensive, we consider enriching the available
labeled data using data augmentation. In a data
augmentation algorithm, an input sentence is perturbed to
create a new sentence that is expected to have the same
meaning or preserving the semantic structure of the
sentence but may accept a slightly difference in lexical
representations. Generally speaking, data augmentation
for text is a hard problem because a small perturbation of
the sentence can create a sentence with a completely new
meaning or with a different semantic structure. Therefore,
we propose a controlled data augmentation to preserve the
main semantic structure of the rules where the spans are
perturbed but the labels are persisted.
For each label v in V, let Span(v) be the collection of all
the spans that are labeled as v in our training data. For
every input labeled sentence among our 141 paragraphs,
we look at each span with label v in the sentence and
randomly replace the given span with a random span
sampled from the set Span(v) to create a new sentence
used for training purposes. This data augmentation
method is simple but it is very effective as demonstrated
in the experiments. Figure 2 shows an example with an
input sentence and a newly generated sentence using this
method. Since the structure of the sentence is preserved,
swapping spans between sentences helps the model
generalize better as it is forced to learn the hidden
structure of the input rather than remembering the actual
span contents.
Model fine-tuning. Our work relies on pretrained
language models such as BERT. Since BERT is trained
on public domain texts, a popular practice when dealing
with domain specific text is to fine-tune these models with
domain specific texts. We consider two approaches to
this. The first one fine-tunes BERT using all policy texts
available in our data. The second method only fine-tunes
BERT on the rules. Figure 3 shows the span prediction F1
scores of different approaches. The combination of
BERT, data augmentation and fine-tuning with rule text
yields the best result so it is the default choice in our
experiments.
Figure 3 F1 score of different approaches for span prediction in a 5-
fold cross-validation settings. The combination of BERT, data aug-
mentation and fine-tuning with rule text yields consistent improvement
over BERT alone.
Ontology-based span prediction to rules
The predicted labels, which capture the structural
information of the rules, are combined with the entity
annotations to construct the benefit rules. For each
sentence, all the labels predicted by the span predictor that
are compatible with a type of benefit rule (e.g., those
corresponding to potential conditions in a Mutually
Exclusive rule as defined in the ontology) are clustered.
Subsequently, we identify for each cluster the entity
annotations that overlap with each label and whose type
is compatible with the range of the predicted label, which
corresponds to a condition defined in the ontology.
At this point, for each label there will be one or more
matching annotations that are used as a value to build the
benefit rule. If there is no match, the text covered by the
label can be used to infer a new entity of the type expected
by the condition label. In our example there are two
clusters, corresponding to a Mutually Exclusive and a
Service Limitation benefit rule. The span predictor
predicts the condition label
“hasMutuallyExclusiveNonReimbursableService” that
covers the text “comprehensive oral evaluation” (Figure
2) and is compatible with a Mutually Exclusive rule. The
ontology annotators identify "d0150 - comprehensive oral
evaluation" as an instance of a “Procedure Code” type and
the annotated span overlaps with the span of the predicted
label. Moreover, the type of the instance is compatible
with the expected range of the condition in the ontology,
therefore, the condition (whose display name in the
ontology is defined as “mutually exclusive - non-
reimbursable service” as shown in Figure 1) is assigned
the instance value “d0150 - comprehensive oral
evaluation” and will be one of the conditions of the built
benefit rule.
In the example, the textual span "seen by a dentist" is
labeled as the non-reimbursable service, while “seen by a
dentist” obviously does not correspond to an actual
instance of a billed service in the ontology, in the context
of this sentence investigators interpret it as "any dental
service". The prediction correctly recognizes that pattern
from similar labeled rules. The extractor can optionally be
configured to allow the creation of benefit rules with
predicted values even if unrecognized in the ontology,
while this may introduce inaccurate results (affecting
precision) it does favour recall. Investigators can then
validate unknown values while validating the benefit rule
and either match them to existent ones (eg., in this case
the service “all dental services”) or optionally update the
ontology with new instances and/or lexicalizations.
Results
Experimental Setup
We evaluate the systems accuracy in terms of
precision/recall when extracting rules from text [5]. The
evaluations are carried out with respect to a gold-standard
of 141 rules created in consultation with our policy
investigators for two US healthcare policies in the dental
domain and compared with respect to the rule-based
approach used as baseline. Given the small number of
training instances, we obtained span prediction for 141
gold sentences in a 5-fold cross-validation setting. This
technique guarantees that predictions are always made
without overlapping between training and test sets. The
span predictions for sentences that do not overlap with the
141 gold sentences (i.e., those that do not contain a benefit
rule and are not filtered out by the classifier) are computed
using a model trained with all the gold sentences. Instead
the span predicton for sentences s that overlap with 141
gold sentences are computed using a model trained
without s.
Evaluation
We compare results with two configurations: default and
optimal. Modifiable configuration includes selecting
whether or not to add values not recognized by the
ontology, choosing to use the classifier, filters, and rule-
based consolidation strategies predefined in the pipeline
[5]. The default configuration was reported in [5], while
the optimal configuration is identified by optimizing for
F1, using the Optuna package [11].
In the first experiment, using the default configuration,
we compare the learning methods with the rule-based
approaches. Table 1 reports the results for both Learn+Uv
and Learn methods, which correspond respectively to the
variant of the method with and without adding unknown
spans to the final benefit rule. We also report results for
execution with (+C) and without the classifier.
In the second experiment we compare the rule-based
method (Rule) with the deep learning approach (Learn),
reporting the maximum performances obtained by both
methods, optimizing the pipeline configuration using the
Optuna package and selecting the best execution from
200 trials in terms of F1 score. This experiment analyzes
the two systems to the best of their performance and
allows comparison of the approaches independently of the
default configuration.
Table 1 Method comparison using the default configuration: Learn-
ing method (Learn) and Rule-based baseline (Rule). (+Uv) Indicates
where the option to keep spans unrecognized in the ontology was used
(+C) Indicated if the classifier was active in the pipeline
Method
Precison
Recall
F1
Learn
69.85
58.23
62.26
Learn+Uv
56.48
69.17
61.53
Learn+C
78.1
54.78
62.82
Learn+C+Uv
67.19
66.36
66.73
Rule
60.57
67
63.34
Rule+C
69.55
63.72
66.47
Table 2 Method comparison using the optimal configuration: Learn-
ing method (Learn). Learning method combined with Rule-based
method (Learn+Rule) and Rule-based method (Rule).
Method
Precison
Recall
F1
Learn
79.56
66
72.03
Rule
76.54
67.37
71.57
Learn+Rule
70.77
73.75
72.22
Discussion and future work
Results reported in Table 1 and Table 2 demonstrate that
the neural approach achieves comparable performance
compared to the rule-based method, learning correlations
from a surprisingly small set of benefit rules provided by
investigators. Table 1 shows the effect of adding or
ignoring values that are not part of the ontology in the
extracted benefit rules and the trade-off between precision
and recall. Intuitively, adding values that are not part of
the ontology increases the recall of the system, extracting
a greater number of rules present in the gold-standard,
while reducing the precision of the system. The classifier,
on the other hand, has the opposite effect, as it increases
precision but reduces recall. This behavior is expected
since the classifier in the pipeline operates as a filter,
removing the paragraphs with low probability of
containing benefit rules.
Table 2 shows the results of various methods in the
optimal configuration environment. The precision of the
learning method is higher than that of the rule-based
method, and the recall values of the two methods are
roughly similar. Although the precision of the Learning
method combined with the Rule-based method has
decreased, the recall has increased significantly from 67%
to 73.75%. The F1 score is 72.22% which shows that the
combination of the Learning method and the Rule-based
method can obtain excellent comprehensive performance.
Furthermore, in Table 1, the use of the span prediction
together with the classifier (Learn+C+Uv) achieved the
best result due to the presence of entities that are not cap-
tured yet in the domain ontology and/or textual patterns
that are notably hard to extract based on pre-defined lin-
guistic and semantic patterns to build a benefit rule. The
result demonstrated that machines are able to learn from
examples of benefit rules validated by our investigators,
without the need of linguistic and semantic experts to add
more hand-crafted linguistic rules or reasoning patterns in
the system [5].
The learning method presented requires an initial
modeling phase, where the main concepts and
relationships are defined and formalized in an ontology.
While many of the concepts are shared and the ontology
contains valid common concepts when applied in the
same domain, as in the case of insurance policy texts of
different US states, in order to apply the method to a new
domain it is necessary to repeat the modeling phase. In
future work we could analyze the use of state-of-art
ontology learning system, such as [12], combined with
the BERT-based extractor to extract rules in a different
domain.
Conclusions
Governments and businesses everywhere are automating
policy enforcement. When they do, the resulting rules
becomes the effective policythat most citizens actually
experience in their lives. That comes with great
opportunities for fairnessboth in enabling policy to be
applied consistently at scale and in defending scarce
resources from wasteful practices not compliant to policy.
This latter point is often missed but it is critical in
ensuring that vulnerable people can access the services
they need. However, organizations that automate policy
enforcement have a responsibility to ensure that there is
no translation of intenterrors when translating from
policy, to business requirements, to rules, to code. A
recent OECD report [3] on this issue identifies several
ways to tackle this, some visionary (simultaneous code
and policy development) and others practical (use AI and
automation to shorten the route from policy to code). Our
system takes this latter approach. This meant developing
a structured representation of the benefit rules that feels
natural and understandable to policy-aware users and is
grounded to specific policy text. This core representation
is captured by the ontology and is critical in ensuring that
non-technical policy-aware individuals can understand
and correct the rules.
In this paper we proposed a learning approach to predict
benefit rule conditions from curated, small labeled
examples that can be added by policy-aware users, and
that can be assembled into actionable benefit rules using
the ontology as a blueprint. We determined that the
learning approach achieves similar performance to a
pattern-based approach. While the training set is currently
small, due to the need for domain experts and the cost of
manual policy labeling, it is expected this will expand as
investigators use the system to review and curate more
benefit rules in any given policy domain.
We believe this work shows the potential impact of lev-
eraging NLP and AI technologies with expert knowledge
in an area where human understanding and control are
important AI design concerns, such as safeguarding the
integrity of government healthcare insurance programs.
This combination empowers investigators to be more
effective and consistent in formalizing and validating
human and machine interpretable rules at scale, and
enables the system to learn from those corrections and
improve its performance over time.
References
[1] J. Li, K.Y. Huang, J. Jin, et al. A survey on statistical meth-
ods for health care fraud detection. Health care management
science (2008), 11(3): 275-87.
[2] H. Joudaki, A. Rashidian, B. Minaei-Bidgoli, et al. Using
data mining to detect health care fraud and abuse: a review
of literature. Global journal of health science (2015),
7(1):194.
[3] J. Mohun, A. Roberts. Cracking the code: Rulemaking for
humans and machines. OECD Working Papers on Public
Governance, No. 42, OECD Publishing (2020)
[4] T.S. Brisimi, V. Lopez, V. Rho, M. Sbodio, G. Picco, M.
Kristiansen, J Segrave-Daly, C Cullen. Ontology-guided
Policy Information Extraction for Healthcare Fraud Detec-
tion. Medical Informatics Europe conference, MIE 2020.
[5] V., Lopez, V., Rho, T.S., Brisimi, F., Cucci, M., Kristian-
sen, J., Segrave-Daly, J., Scalvini, J., Davis, G., Ferguson.
Benefit graph extraction from healthcare policies. Interna-
tional Semantic Web Conference, ISWC 2019
[6] https://cloud.ibm.com/docs/compare-comply
[7] Iowa Department of Human Services (DHS) Dental Ser-
vices Provider Manual (2013) Retrieved from
https://dhs.iowa.gov/sites/default/files/Den-
tal.pdf?012520210943
[8] Centers for Medicare & Medicaid Services (U.S.)
Healthcare Common Procedure Coding System (HCPCS).
Centers for Medicare & Medicaid Services (2020).
[9] Centers for Medicare & Medicaid Services (U.S.). n.d.
Place of Service Code (2020). https://www.cms.gov/Medi-
care/Coding/place-of-service-codes
[10] J. Devlin, M.W. Chang, K. Lee, and K. Toutanova. 2019.
BERT: Pre-training of deep bidirectional transformers for
language understanding. ACL
[11] T. Akiba, S. Sano, T. Yanase, T. Ohta, and M. Koyama.
Optuna: A Next-generation Hyperparameter Optimization
Framework SIGKDD 2019.
[12] A. Khadir, H. Aliane and AGuessoum. 2021. Ontology
learning: Grand tour and challenges. Computer Science Re-
view. 39. 100339. 10.1016/j.cosrev.2020.100339.
[13] Kingma, Diederik P., and Jimmy Ba. "Adam: A
method for stochastic optimization." arXiv preprint
arXiv:1412.6980 (2014).
Address for correspondence
Gabriele Picco
IBM Reseach Europe, Dublin, Ireland
Email address: gabriele.picco@ibm.com
ResearchGate has not been able to resolve any citations for this publication.
Article
Full-text available
Financial losses in Medicaid, from Fraud, Waste and Abuse (FWA), in the United States are estimated to be in the tens of billions of dollars each year. This results in escalating costs as well as limiting the funding available to worthy recipients of healthcare. The Centers for Medicare & Medicaid Services mandate thorough auditing, in which policy investigators manually research and interpret the policy to validate the integrity of claims submitted by providers for reimbursement, a very time-consuming process. We propose a system that aims to interpret unstructured policy text to semi-automatically audit provider claims. Guided by a domain ontology, our system extracts entities and relations to build benefit rules that can be executed on top of claims to identify improper payments, and often in turn payment policy or claims adjudication system vulnerabilities. We validate the automatic knowledge extraction from policies based on ground truth created by domain experts. Lastly, we discuss how the system can co-reason with human investigators in order to increase thoroughness and consistency in the review of claims and policy, to identify providers that systematically violate policies and to help in prioritising investigations.
Conference Paper
Full-text available
With healthcare fraud accounting for financial losses of billions of dollars each year in the United States, the task of investigating regulation adherence is key to reduce the impact of Fraud, Waste and Abuse (FWA) on the healthcare industry. Providers rendering services to patients typically submit claims to healthcare insurance agencies. Such claims must follow specific compliance criteria specified by state and federal policies. This paper presents an ontology-based system that aims to support the FWA claim investigation process by extracting graph-based actionable knowledge from policy text de-scribing those compliance criteria. We discuss the process of creating a do-main-specific ontology to model human experts’ conceptualisations and to incorporate early-on the feedback of FWA investigators, who are the early adopters of our solution. We explore whether the ontology is expressive and flexible enough to model the diverse compliance processes and complex rela-tionships defined in policy documents. The ontology is then used, in combination with natural language understanding and semantic techniques, to guide the extraction of a Knowledge Graph (KG) from policies. Our solution is validated in terms of correctness and completeness by comparing the extracted knowledge to a ground truth created by investigators. Lastly, we discuss further challenges our deployed semantic system needs to tackle in this novel scenario, with the prospect of supporting the investigation process.
Article
Full-text available
Inappropriate payments by insurance organizations or third party payers occur because of errors, abuse and fraud. The scale of this problem is large enough to make it a priority issue for health systems. Traditional methods of detecting health care fraud and abuse are time-consuming and inefficient. Combining automated methods and statistical knowledge lead to the emergence of a new interdisciplinary branch of science that is named Knowledge Discovery from Databases (KDD). Data mining is a core of the KDD process. Data mining can help third-party payers such as health insurance organizations to extract useful information from thousands of claims and identify a smaller subset of the claims or claimants for further assessment. We reviewed studies that performed data mining techniques for detecting health care fraud and abuse, using supervised and unsupervised data mining approaches. Most available studies have focused on algorithmic data mining without an emphasis on or application to fraud detection efforts in the context of health service provision or health insurance policy. More studies are needed to connect sound and evidence-based diagnosis and treatment approaches toward fraudulent or abusive behaviors. Ultimately, based on available studies, we recommend seven general steps to data mining of health care claims.
Article
Full-text available
Fraud and abuse have led to significant additional expense in the health care system of the United States. This paper aims to provide a comprehensive survey of the statistical methods applied to health care fraud detection, with focuses on classifying fraudulent behaviors, identifying the major sources and characteristics of the data based on which fraud detection has been conducted, discussing the key steps in data preprocessing, as well as summarizing, categorizing, and comparing statistical fraud detection methods. Based on this survey, some discussion is provided about what has been lacking or under-addressed in the existing research, with the purpose of pinpointing some future research directions.
Conference Paper
The purpose of this study is to introduce new design-criteria for next-generation hyperparameter optimization software. The criteria we propose include (1) define-by-run API that allows users to construct the parameter search space dynamically, (2) efficient implementation of both searching and pruning strategies, and (3) easy-to-setup, versatile architecture that can be deployed for various purposes, ranging from scalable distributed computing to light-weight experiment conducted via interactive interface. In order to prove our point, we will introduce Optuna, an optimization software which is a culmination of our effort in the development of a next generation optimization software. As an optimization software designed with define-by-run principle, Optuna is particularly the first of its kind. We will present the design-techniques that became necessary in the development of the software that meets the above criteria, and demonstrate the power of our new design through experimental results and real world applications. Our software is available under the MIT license (https://github.com/pfnet/optuna/).
Cracking the code: Rulemaking for humans and machines
  • J Mohun
  • A Roberts
J. Mohun, A. Roberts. Cracking the code: Rulemaking for humans and machines. OECD Working Papers on Public Governance, No. 42, OECD Publishing (2020)
) Healthcare Common Procedure Coding System (HCPCS)
Centers for Medicare & Medicaid Services (U.S.) Healthcare Common Procedure Coding System (HCPCS). Centers for Medicare & Medicaid Services (2020).
BERT: Pre-training of deep bidirectional transformers for language understanding
  • J Devlin
  • M W Chang
  • K Lee
  • K Toutanova
J. Devlin, M.W. Chang, K. Lee, and K. Toutanova. 2019. BERT: Pre-training of deep bidirectional transformers for language understanding. ACL
Ontology learning: Grand tour and challenges. Computer Science Review
  • A Khadir
  • H Aliane
A. Khadir, H. Aliane and AGuessoum. 2021. Ontology learning: Grand tour and challenges. Computer Science Review. 39. 100339. 10.1016/j.cosrev.2020.100339.