Information-Theoretic Patient Record Matching in Medical
Databases: A Discriminative Power and Feature Analysis Using
MIMIC-IV
Vitalijs Teze
a
, Erika Nazaruka
b
and Dmirtijs Bliznuks
c
Institute of Applied Computer Systems, Riga Technical University, 10 Zunda Embankment, Riga, Latvia
Keywords: Patient Record Matching, Information Theory, Feature Extraction, MIMIC-IV, Critical Care, Shannon
Entropy, Feature Stability, Patient Identification, Healthcare Data Integration, Temporal Pattern Analysis.
Abstract: This paper presents an information-theoretic framework to evaluate feature discriminative power and stability
for patient record matching. We analyse the discriminative power and temporal stability of features through
Shannon entropy, evaluating their effectiveness for patient identification without unique identifiers. Our
framework categorizes features into demographics/administrative (𝐷
(
𝐹
)
=12247.56 bits), ICU care patterns
(𝐷
(
𝐹
)
=266.40 bits), and clinical records (𝐷
(
𝐹
)
=12.10 bits), achieving a combined discriminative power of
12526.06 bits. This significantly exceeds the theoretical minimum threshold (log

(
𝑁
)
≈16 bits) for our
population of 65,366 patients. The framework employs hierarchical feature weighting based on information
content and stability coefficients, revealing that temporal patterns and service transitions contain higher
discriminative power than traditional demographic identifiers. We demonstrate that effective matching
requires balancing feature stability against information content while maintaining computational efficiency.
The framework provides a foundation for implementing reliable patient matching systems, though further
validation across diverse healthcare environments is needed.
1 INTRODUCTION
Healthcare data is recognized as a cornerstone for
improving patient outcomes, optimizing resource
utilization, and advancing medical research. In
critical care settings, the ability to accurately link
patient records across fragmented datasets is essential
for ensuring continuity of care and conducting robust
retrospective analyses (Duggal et al., 2015; Kho et al.,
2015). In a fragmented healthcare system, such as in
the United States, where patients may receive care at
multiple institutions, individual-level data is often
scattered across disparate systems (Ong et al., 2020).
Accurately linking patient records is crucial for
creating a comprehensive view of a patient's medical
history and enabling data analysis across a wider
range of inquiries pertinent to research, clinical, and
commissioning arenas (Clark et al., 2019).
a
https://orcid.org/0009-0001-2165-4789
b
https://orcid.org/0000-0002-1731-989X
c
https://orcid.org/0000-0003-4252-9220
However, achieving reliable record matching is a
persistent challenge in the absence of unique patient
identifiers (Duggal et al., 2015; Kho et al., 2015; Ong
et al., 2020). Without a nationwide unique patient
identifier, accurately matching multiple records for
the same patient from disparate sources is
challenging, particularly in large and complex
datasets (Duggal et al., 2015). The reliance on basic
matching methods using existing records often results
in inaccurate patient identification (Fernandes &
O’Connor, 2015).
Patient record matching typically relies on
identifiers such as social security numbers or medical
record numbers (Duggal et al., 2015; Godlove & Ball,
2015). When these are unavailable, alternative
approaches must leverage features such as
demographics, clinical events, and temporal patterns
(Evans et al., 2016; Nie & Roantree, 2019). These
approaches may include deterministic and
probabilistic matching algorithms. Deterministic
280
Teze, V., Nazaruka, E. and Bliznuks, D.
Information-Theoretic Patient Record Matching in Medical Databases: A Discriminative Power and Feature Analysis Using MIMIC-IV.
DOI: 10.5220/0013475400003928
Paper published under CC license (CC BY-NC-ND 4.0)
In Proceedings of the 20th International Conference on Evaluation of Novel Approaches to Software Engineering (ENASE 2025), pages 280-291
ISBN: 978-989-758-742-9; ISSN: 2184-4895
Proceedings Copyright © 2025 by SCITEPRESS – Science and Technology Publications, Lda.

matching relies on exact matches of key identifiers,
while probabilistic matching uses algorithms to
calculate the likelihood of a match based on the
similarity of data elements (Blecker et al., 2016;
Godlove & Ball, 2015; McCoy et al., 2013; Riplinger
et al., 2020).
Despite the availability of sophisticated methods,
existing research often lacks a systematic approach to
identifying the minimal set of features necessary for
accurate matching, particularly in critical care
contexts where data heterogeneity and missing values
are prevalent. The reliance on a limited number of
identifiers and inconsistencies in data quality often
lead to suboptimal matching rates. Standardizing
demographic data elements, such as telephone
numbers, dates of birth, and addresses, can improve
matching algorithm accuracy (Godlove & Ball, 2015;
Riplinger et al., 2020).
This paper addresses this gap by presenting a
probabilistic matching framework tailored for the
Medical Information Mart for Intensive Care
(MIMIC)-IV database. Our framework uses Shannon
entropy to quantify feature utility, distinct from
deployable matching algorithms. Our primary focus
is to determine the minimal set of features required
for reliable patient matching and to evaluate the
performance of our approach under realistic clinical
scenarios. By doing so, we aim to provide a scalable
and practical solution for healthcare data integration
in critical care settings. Evaluating the effectiveness
of referential matching software, which augments
patient data with information from external sources,
and exploring big data analytics approaches like
fuzzy matching algorithms and MapReduce
techniques can potentially enhance matching rates
and improve clinical decision-making.
This study addresses two questions: (1) Which
features provide the greatest discriminative power for
patient matching in MIMIC-IV? (2) How do stability
and information content trade off in feature selection?
2 RELATED WORK
The primary challenge lies in accurately linking
records from disparate sources that pertain to the
same individual without compromising patient
privacy. Inaccurate matching can lead to medical
errors, compromised patient safety, billing mistakes,
and flawed research outcomes (Fernandes &
O’Connor, 2015; Godlove & Ball, 2015; Just et al.,
2016; Riplinger et al., 2020; Zech et al., 2016).
2.1 Matching Techniques
Unique Patient Identifiers (UPIs): Some countries,
such as Singapore, Canada, and Australia, have
implemented national healthcare identifiers to
facilitate patient matching. However, these identifiers
often face limitations in cross-border information
sharing and incorporating data from non-traditional
sources like social care settings (Fernandes &
O’Connor, 2015). In the United States, the lack of a
nationwide unique patient identifier poses a
significant challenge to accurately matching records
(Duggal et al., 2015; Godlove & Ball, 2015).
Algorithmic Approaches: These methods utilize
demographic data, such as name, date of birth, social
security number, and address, to match patient
records. Algorithms range in complexity from basic
deterministic matching, requiring exact matches on
specific identifiers, to sophisticated probabilistic
matching techniques that employ statistical models
and threshold limits. However, even advanced
algorithms fall short of achieving a 100% match rate
(Fernandes & O’Connor, 2015; Riplinger et al.,
2020).
Referential Matching Software. This approach
enhances algorithmic matching by utilizing third-
party databases containing verified patient
information. This supplementary data can help
resolve ambiguities and improve match rates
(Riplinger et al., 2020).
Hybrid Models. Recognizing the limitations of
individual approaches, researchers have proposed
hybrid models that combine different techniques. For
example, combining algorithmic matching with
referential matching software can potentially enhance
accuracy. Other examples include combining
structured and unstructured data. Big data analytics
techniques like fuzzy matching algorithms and
MapReduce have also been proposed for handling
large datasets (Blecker et al., 2016; Duggal et al.,
2015; Riplinger et al., 2020).
Privacy-Preserving Record Linkage (PRL). With
growing concerns about patient privacy, researchers
are actively developing techniques that enable record
linkage without disclosing sensitive patient
identifiers. These techniques often involve masking
or encrypting identifiers before performing matching
operations. Examples of such techniques include
using Bloom filters and one-way hashing algorithms
to protect patient privacy while enabling record
linkage (Godlove & Ball, 2015; Sehili et al., 2015;
Toth et al., 2014; Vatsalan et al., 2017).
Machine Learning and Deep Learning Methods:
Recent advancements in machine learning have
Information-Theoretic Patient Record Matching in Medical Databases: A Discriminative Power and Feature Analysis Using MIMIC-IV
281

significantly improved patient identification and
matching in healthcare by leveraging both structured
and unstructured data. Integrating data types such as
demographics, clinical notes, and diagnostic codes
allows these algorithms to capture complex patterns
and nuances that traditional rule-based methods often
miss. Deep learning models have shown remarkable
potential in enhancing accuracy and inclusivity,
especially when applied to diverse patient
populations and challenging clinical scenarios. The
approach might overcome limitations of rule-based
systems, although it is constrained by dataset
diversity (Blecker et al., 2016; Gehrmann et al., 2018;
Hua et al., 2023).
2.2 Gap Analysis
Despite the breadth of existing research on patient
matching — ranging from algorithmic approaches to
referential matching and hybrid models — none of
the works discussed above explicitly apply
Information Theory (Shannon, 1948) in evaluating or
optimizing patient record linkage. Scholars often
focus on improving matching accuracy via
algorithmic refinements (deterministic, probabilistic,
referential), but do not frame the problem in terms of
entropy or minimal information requirements.
In applying an information-theoretic viewpoint,
we aim to systematically quantify how much
information each feature (or combination of features)
carries by computing their entropy, assessing joint
and conditional entropies, and comparing the
cumulative information gained to log

(N) — the
theoretical threshold for uniquely identifying a single
patient within a population of size N . We will
demonstrate how this approach can be applied to the
MIMIC-IV database, leveraging its set of attributes
and patient cohort to evaluate the minimal set of
features required for both reliable and efficient patient
matching.
3 METHODOLOGY
3.1 Data Description
We utilized the MIMIC-IV database (v3.1), which
contains de-identified electronic health record data of
patients admitted to the Beth Israel Deaconess
Medical Center in Boston, MA. The dataset spans
from 2008 to 2022 (Goldberger et al., 2000; A.
Johnson et al., 2024; A. E. W. Johnson et al., 2023).
Our analysis focuses on Intensive Care Unit (ICU)
stays, resulting in a subset of 65366 unique patient
admissions.
There are 31 tables from 2 schemas (hosp, icu) of
the MIMIC-IV database that we have access to.
Several auxiliary tables in MIMIC-IV contain only
identifier mappings without additional attributes that
could contribute to patient matching. For example,
the caregiver and provider tables consist solely of
unique identifiers (caregiver_id and provider_id
respectively) without any supplementary information
about the healthcare providers themselves. While
these tables are essential for maintaining referential
integrity within the database, they do not provide
discriminative features for patient identity resolution
and are therefore excluded from our analysis.
The MIMIC-IV database structure for patient
record matching can be organized into four main
categories:
1. Core Patient Data
 Patient demographics (patients table): Contains
fundamental identifiers (subject_id),
demographics, and mortality data.
 Hospital encounters (admissions, transfers,
icustays tables): Track patient movement
through hospital units using hadm_id and
stay_id as linking keys.
 Clinical assignments (services table):
Documents care team responsibilities
independent of physical location.
 Outpatient measurements (omr table): Contains
longitudinal measurements like vital signs and
anthropometrics.
2. Clinical Events and Orders
 ICU documentation (chartevents): Main
repository (313M+ rows) for vital signs, labs,
and clinical measurements.
 Temporal events (datetimeevents,
ingredientevents, inputevents, outputevents):
Track time-based clinical activities and patient
I/O.
 Procedures and diagnoses: Captured through
multiple coding systems (ICD-9/10, CPT) in
diagnoses_icd, procedures_icd, drgcodes, and
hcpcsevents tables.
 Provider orders (poe, poe_detail):
Comprehensive order tracking using a flexible
EAV model.
3. Laboratory and Diagnostic Data
 Laboratory measurements (labevents):
Specimen-linked test results with 98% hospital
stay coverage.
ENASE 2025 - 20th International Conference on Evaluation of Novel Approaches to Software Engineering
282

 Microbiology cultures (microbiologyevents):
Hierarchical culture results including negative
findings and antibiotic sensitivities.
4. Medications and Prescriptions
 Administration records (emar, emar_detail):
Barcode-verified medication delivery data
post-2011.
 Pharmacy management (pharmacy,
prescriptions): Detailed medication orders with
standardized identifiers (GSN, NDC).
All tables are interconnected through key
identifiers (subject_id, hadm_id, stay_id) and
supported by reference tables (d_icd_diagnoses,
d_icd_procedures, d_labitems, d_items, d_hcpcs)
that provide standardized definitions and
classifications.
3.2 Feature Extraction Framework
In patient record matching without unique identifiers,
categorical variables play a critical role in
establishing identity linkage. However, the reliability
and discriminative power of these variables depends
heavily on their consistency across multiple
encounters. We evaluated categorical features across
the MIMIC-IV database to assess their suitability for
patient matching.
First, we identified potentially useful categorical
variables across major database tables, focusing on
features that could contribute to patient identification.
Demographic Identifiers: Gender, race, language,
and marital status from patients and admissions
tables; Insurance type and admission location;
Admission type (including AMBULATORY
OBSERVATION, DIRECT EMER., ELECTIVE,
etc.); Anchor year grouping formatted as specific year
ranges (e.g., 2008-2010).
Clinical Service Patterns: Care unit transitions;
Clinical services (e.g., CMED, CSURG, DENT);
Event types for transfers (ed, admit, transfer,
discharge); Admission and transfer patterns.
Clinical Categorizations: Laboratory test priorities
(routine/stat) and flags; Medication routes and types
(MAIN, BASE, ADDITIVE); Order types and status
(Active/Inactive); Procedure status (Paused,
FinishedRunning, Stopped); Administration types
from medication records; Specimen types from
microbiology data.
For each categorical variable, we analyzed the
following metrics:
 The number of subjects with multiple distinct
values.
 The percentage of subjects showing variation
across encounters.
 The total number of unique values in each
category.
 The distribution of values across the
population.
This analysis serves multiple purposes:
 Information Content Assessment: Variables
with high consistency across encounters but
good variation across the population provide
strong discriminative power.
 Data Quality Evaluation: Unexpected
variations in supposedly stable characteristics
(e.g., multiple recorded genders) may indicate
data quality issues.
 Feature Selection: Guides the selection of
reliable categorical features for the matching
algorithm.
 Entropy Calculation: Informs the theoretical
information content available from each
categorical variable.
In addition to categorical variables, temporal
patterns provide important discriminative
information for patient matching. We analyzed
temporal features across multiple dimensions:
Admission Patterns: Time intervals between
hospital admissions; Emergency department
registration to admission delays; Length of stay
distributions; Season and time-of-day admission
patterns.
Care Transitions: Service-to-service transfer
intervals; ICU transfer timing sequences;
Department-to-department movement patterns.
Treatment Timelines: Laboratory test ordering
patterns; Medication administration sequences (from
emar); Procedure scheduling patterns.
Documentation Patterns: Time deltas between chart
time and store time across various events (lab results,
medications, procedures); Order-to-administration
intervals.
For each temporal feature, we computed:
 Intra-patient timing consistency.
 Inter-patient timing variations.
 Cyclic pattern detection (daily, weekly,
seasonal).

Sequential pattern stability.
The temporal analysis provides:
 Pattern Recognition: Identification of
characteristic temporal signatures in patient
care sequences.
Information-Theoretic Patient Record Matching in Medical Databases: A Discriminative Power and Feature Analysis Using MIMIC-IV
283

 Timing Fingerprints: Development of patient-
specific temporal patterns that persist across
encounters.
 Quality Control: Detection of temporal
anomalies that might indicate record matching
errors.
 Information Gain: Quantification of additional
discriminative power when temporal features
are combined with categorical variables.
This temporal dimension adds context to our
matching framework, particularly for distinguishing
between patients with similar categorical profiles but
distinct care patterns.
3.3 Feature Analysis Framework
Our analysis framework employs information theory
to evaluate the discriminative power and reliability of
features for patient matching. This approach consists
of three main components:
3.3.1 Information Content Analysis
For each feature f, we compute:
Shannon entropy (1), where 𝑝(𝑥) is the
probability of value 𝑥.
𝐻
(
𝑓
)
=−𝑝
(
𝑥
)
log

𝑝
(
𝑥
)
(1)
Conditional entropy (2) for feature 𝑓 given
subject 𝑠.
𝐻
(
𝑓
|
𝑠
)
=−𝑝
(
𝑠
)
𝑝
(
𝑥
|
𝑠
)

log

𝑝
(
𝑥
|
𝑠
)

(2)
Mutual information (3) between feature and
subject identity.
𝐼
(
𝑓
;𝑠
)
=𝐻
(
𝑓
)
−𝐻
(
𝑓
|
𝑠
)
(3)
Null rate penalty factor (4).
α
(
𝑓
)
=1−
null_count
total_records
(4)
The effective information content 𝐼

(
𝑓
)
is
calculated as the product of mutual information and
the null rate penalty factor (5).
𝐼

(
𝑓
)
=𝐼
(
𝑓
;𝑠
)
×α
(
𝑓
)
(5)
3.3.2 Temporal Stability Assessment
For features that vary over time, we evaluate:
Intra-patient variance (6) computed across
multiple encounters. Where 𝑁 is the number of
patients, 𝑛

is the number of encounters for patient 𝑖,
𝑓

is the feature value for patient 𝑖 at encounter 𝑗, 𝑓


(7) is the mean feature value for patient 𝑖.
σ


(
𝑓
)
=
1
𝑁

1
𝑛

−1

𝑓
,
−
𝑓










(6)
𝑓


=
1
𝑛


𝑓
,




(7)
Inter-patient variance (8) between different
patients. Where 𝑁 is the number of patients, 𝑓


(7) is
the mean feature value for patient 𝑖, 𝑓
̅
(9) is the grand
mean across all patients.
σ


(
𝑓
)
=
1
𝑁−1

𝑓


−
𝑓
̅





(8)
𝑓
̅
=
1
𝑁

𝑓





(9)
Stability coefficient (10).
𝑆
(
𝑓
)
=
σ


(
𝑓
)
σ



(
𝑓
)
(10)
Time-decay factor (11) modelling feature
stability over temporal gaps. Where 𝛥𝑡 is the time
difference between measurements, 𝛼

(12) is the
decay constant specific to feature 𝑓, 𝑡
/
(
𝑓
)
is the
half-life period for feature 𝑓, 𝑠(𝑓) (13) is the feature
stability score, 𝐼(𝑐𝑜𝑛𝑑𝑖𝑡𝑖𝑜𝑛) is the indicator function
(1 if true, 0 if false), 𝑁 is the number of patients, 𝑛

is
the number of encounters for patient 𝑖, 𝑓
,
is the
feature value for patient 𝑖 at encounter 𝑗.
The half-life 𝑡
/
(
𝑓
)
can be empirically
determined for each feature type. For example:
 Demographics (gender, race): Very long half-
life (years).
 Insurance status: Medium half-life (months).
 Clinical measurements: Short half-life
(days/weeks).
λ
(
f
,Δt
)
=e



×S
(
f
)
(11)
ENASE 2025 - 20th International Conference on Evaluation of Novel Approaches to Software Engineering
284

α

=−
ln
(
0.5
)
𝑡
/
(
𝑓
)
(12)
𝑠
(
𝑓
)
=
1
𝑁

𝐼
𝑓
,
=
𝑓
,

𝑛

−1







(13)
3.3.3 Hierarchical Feature Weighting
Two stability thresholds are defined for feature
classification:
 threshold₁ = 0.3: Upper bound for high stability
features
 threshold₂ = 0.7: Upper bound for moderate
stability features
Thresholds (0.3, 0.7) were set at the 25th and 75th
percentiles of stability coefficients across MIMIC-IV
features, balancing high-stability primaries (e.g.,
gender) and variable tertiaries (e.g., ICU duration).
Features are classified into tiers based on their
information content and stability:
Primary Features (Tier 1, 𝛽
tier
=1):
 High stability. 𝑆(𝑓) <= threshold₁
 Low entropy 𝐻(𝑓) <= 2
 Minimal null rate (𝛼(𝑓) > 0.99)
Secondary Features (Tier 2, 𝛽
tier
=0.5):
 Moderate stability. threshold₁ < 𝑆(𝑓) <=
threshold₂
 Medium entropy 2<𝐻(𝑓) <= 12
 Acceptable null rate. 𝛼(𝑓) > 0.95
3. Tertiary Features (Tier 3, 𝛽
tier
= 0.25):
 Variable stability.
 High entropy.
 Used for disambiguation.
The final feature weight is computed as (14),
where 𝛽
tier
is a tier-specific base weight.
𝑤
(
𝑓
)
=β
tier
×𝐼

(
𝑓
)
×
1
𝑆
(
𝑓
)
(14)
3.3.4 Combined Feature Space
The total discriminative power 𝐷 of a feature set 𝐹 is
evaluated against the theoretical minimum required
information content (15) for 𝑓 ∈ 𝐹.
𝐷
(
𝐹
)
=𝑤
(
𝑓
)

∈
×𝐼

(
𝑓
)
(15)
This must satisfy (16) where N is the total patient
population size.
𝐷
(
𝐹
)
≥log

(
𝑁
)
(16)
4 RESULTS AND DISCUSSION
Our analysis of the MIMIC-IV database (v3.1)
examined the information content and discriminative
power of various features for patient matching. The
study encompassed core hospital data, admission
records, intensive care unit information, and
outpatient measurements from the Online Medical
Record (OMR) system.
The analysed features fall into six distinct
categories: Demographics, Admission Patterns,
Clinical Services, ICU Stays, Care Transitions, and
Outpatient Measurements. Initial analysis covered 27
distinct features across these categories, with results
indicating that temporal patterns and service
transitions contain significantly higher information
content than demographic data alone.
The information theoretic analysis showed that:
 Only Online Medical Record sequential
measurements exceeded the theoretical
minimum information threshold (log₂(N) ≈ 16
bits for our population of N=65,366 patients).
 A cluster of high-information features (service
transitions, ICU stays, care transitions)
provided between 12-16 bits of information,
followed by a significant gap to the next feature
at 6.107 bits.
 Clinical time-based features consistently
showed higher discriminative power than static
demographic data, with traditional identifiers
providing less than 2 bits of information.
4.1 Categorical Feature Analysis
Based on the framework defined in section 3.3.1, we
analyzed each feature's information content through
four metrics:
 Shannon entropy (
𝐻
(
𝑓
)
),
 Conditional entropy ( 𝐻
(
𝑓
|
𝑠
)
),
 Mutual information ( 𝐼
(
𝑓;𝑠
)
),
 The null rate penalty factor (
α
(
𝑓
)
).
The effective information content 𝐼

(
𝑓
)
was then
calculated as the product of mutual information and
the null rate penalty factor.
Table 1 presents the information theoretic metrics
for the categorical features analyzed, including
Shannon
entropy
(
𝐻
(
𝑓
) )
,
conditional entropy
Information-Theoretic Patient Record Matching in Medical Databases: A Discriminative Power and Feature Analysis Using MIMIC-IV
285

Table 1: Information theoretic metrics for the main categorical features.
Cate
g
or
y
Feature H
(
f
)
H
(
f
|
s
)
I
(
f;s
)
I
E
(
f
)
Demo
g
ra
p
hics
g
ende
r
0.998 0.000 0.998 0.998
Demographics anchor_age 6.107 0.000 6.107 6.107
Demographics anchor_year_group 2.283 0.000 2.283 2.283
Admission Patterns race 2.361 0.100 2.262 2.262
Admission Patterns language 0.765 0.000 0.765 0.765
Admission Patterns marital_status 1.680 0.077 1.602 1.602
Admission Patterns insurance 1.649 0.168 1.481 1.481
Admission Patterns admission_type 2.644 0.959 1.685 1.685
Admission Patterns admission_location 2.112 0.742 1.370 1.370
Admission Patterns discharge_location 2.114 0.648 1.466 1.466
Care Transitions careunit 3.599 2.046 1.554 1.554
Care Transitions eventtype 1.980 1.653 0.327 0.327
Care Transitions unit_stay_duration 15.720 3.319 12.401 12.401
Care Transitions transfer_patterns 18.274 3.930 14.344 14.344
Clinical Services curr_service 2.840 0.732 2.108 2.108
Clinical Services
p
rev_service 2.962 0.546 2.416 2.416
Clinical Services service_transition_timing 18.440 2.514 15.926 15.926
ICU Stays first_careunit 2.908 0.399 2.509 2.509
ICU Stays last_careunit 2.908 0.399 2.509 2.509
ICU Stays los 16.317 0.797 15.520 15.520
ICU Stays icu_duration 16.317 0.797 15.520 15.520
ICU Stays unit_transitions 2.908 0.399 2.509 2.509
ICU Stays readmission_intervals 14.824 1.234 13.590 13.590
Online Medical Recor
d
result_name 1.958 1.822 0.135 0.135
Online Medical Recor
d
measurement_patterns 3.473 2.459 1.014 1.014
Online Medical Recor
d
value_patterns 11.417 5.993 5.423 5.423
Online Medical Recor
d
sequential_measurements 17.563 0.001 17.562 17.562
Note: H(f) = Shannon entropy, H(f|s) = Conditional entropy, I(f;s) = Mutual information, I
E
(f) = Effective information
content. All values in bits.
(
𝐻
(
𝑓
|
𝑠
) )
, and mutual information
(
𝐼
(
𝑓;𝑠
) )
. All
features exhibited complete data availability (null rate
penalty factor α
(
𝑓
)
=1.0), indicating no missing
values in the examined MIMIC-IV subset—or an
oversight in detecting them—thus making effective
information content
(
𝐼

(
𝑓
) )
equivalent to mutual
information
(
𝐼
(
𝑓;𝑠
) )
.
The analysis revealed distinct tiers of feature
informativeness:
High Information Content (>12 bits):
 Online Medical Record sequential
measurements (17.562 bits).
 Clinical service transition timing (15.926 bits).
 ICU length of stay and duration (15.520 bits
each) - notably, while 'los' was provided
directly in MIMIC-IV, 'duration' was
calculated as the difference between ‘intime’
and ‘outtime’; their identical information
content validates data consistency.
 Care transition patterns (14.344 bits).
 Readmission intervals (13.590 bits).
 Unit stay duration (12.401 bits).
Medium Information Content (2-6 bits):
 Anchor age (6.107 bits).
 Online Medical Record value patterns (5.423
bits).
 ICU unit features (2.509 bits).
 Service assignments (2.108-2.416 bits).
Low Information Content (<2 bits):
 Traditional demographics (gender: 0.998 bits,
language: 0.765 bits)
 Basic admission data (1.370-1.685 bits)
 Event types (0.327 bits)
 Result names (0.135 bits)
ENASE 2025 - 20th International Conference on Evaluation of Novel Approaches to Software Engineering
286

Table 2: Temporal stability analysis results.
Cate
g
or
y
Feature σ²ᵢₙ
ₜ
ᵣₐ
(
f
)
σ²ᵢₙₜₑᵣ
(
f
)
S
(
f
)
λ
(
f,Δt
)
s
(
f
)
Demo
g
ra
p
hics
g
ende
r
0.000 0.249 0.000 0.000 1.000
Demo
g
ra
p
hics anchor
_
a
g
e 1.248E-33 0.076836395 1.624E-32 1.620E-32 1.000
Demo
g
ra
p
hics anchor
_y
ear
_g
rou
p
0.000 2.069 0.000 0.000 1.000
Admission Patterns race 1.225E+36 9.108E+36 0.134 0.134 0.940
Admission Patterns lan
g
ua
g
e 36900.595 2.924E+36 1.262E-32 1.262E-32 1.000
Admission Patterns marital
_
status 1.171E+36 1.196E+37 0.098 0.096 0.965
Admission Patterns insurance 6.667E+36 4.354E+37 0.153 0.150 0.915
Admission Patterns admission
_
t
yp
e 1.379E+37 1.420E+37 0.971 0.586 0.339
Admission Patterns admission
_
location 2.402E+37 2.511E+37 0.957 0.438 0.478
Admission Patterns dischar
g
e
_
location 1.186E+37 1.637E+37 0.725 0.319 0.551
Care Transitions careunit 1.607E+37 4.733E+36 3.396 2.744 0.230
Care Transitions eventtype 1.702E+37 5.677E+36 2.999 2.297 0.255
Care Transitions unit_stay_duration 5239.687 1237.562 4.234 3.246 0.000
Care Transitions transfer_patterns 101250993.455 79868037.776 1.268 1.009 1.668E-05
Clinical Services curr_service 1.195E+37 1.446E+37 0.826 0.464 0.462
Clinical Services
p
rev_service 1.830E+37 1.497E+37 1.222 0.848 0.151
Clinical Services service_transition_timing 258013226.523 277993281.699 0.928 0.493 0.000
ICU Stays first_careunit 2.161E+37 3.489E+37 0.619 0.277 0.459
ICU Stays last_careunit 2.161E+37 3.489E+37 0.619 0.278 0.459
ICU Stays los 30.752 21.453 1.433 0.728 0.000
ICU Stays icu_duration 17713.190 12356.699 1.433 0.728 0.000
ICU Stays unit_transitions 9.853E+36 1.383E+37 0.712 0.318 0.459
ICU Stays readmission_intervals 212203693.193 239175517.099 0.887 0.564 0.000
Online Medical
Recor
d
result
_
name 3.037E+37 6.653E+36 4.565 3.680 0.166
Online Medical
Recor
d
measurement_patterns 12397959.817 3136807.982 3.952 2.979 0.443
Online Medical
Recor
d
value
_p
atterns 2.848E+37 5.460E+36 5.216 4.205 0.014
Online Medical
Recor
d
sequential_measurements 9.985E+35 1.047E+35 9.541 7.188 0.957
Note: σ²ᵢₙₜᵣₐ(f)= Intra-patient Variance, σ²ᵢₙₜₑᵣ(f) = Inter-patient Variance, S(f) = Stability Coefficient, λ(f,Δt)= Time Decay,
s(f)
=
f
eature stabilit
y
score.
Notably, conditional entropy values revealed that
temporal and sequential features (like Online Medical
Record measurements and service transitions)
retained significant information content even after
accounting for patient identity, indicating their value
for disambiguation.
4.2 Temporal Feature Analysis
Analysis of temporal stability metrics across feature
categories revealed distinct patterns in feature
reliability and degradation over time. Table 2 presents
the complete temporal analysis results.
Key findings include:
Demographics and Persistent Features
(Stability Score (𝑠(𝑓) )> 0.90):
 Basic demographic features showed perfect
stability (gender: 1.0, language: 1.0)
 Race and marital status maintained high
stability (0.94 and 0.96 respectively)
 Insurance information showed good stability
(0.92)
Clinical Service Data (𝑠(𝑓) 0.40-0.50):
 First/last careunit: 0.46 stability
 Current service: 0.46 stability
 Admission location: 0.48 stability
 Service transitions showed moderate stability
with temporal variability
Information-Theoretic Patient Record Matching in Medical Databases: A Discriminative Power and Feature Analysis Using MIMIC-IV
287

Dynamic Care Features (𝑠(𝑓) 0.20-0.35):
 Care unit events: 0.23 stability
 Event type patterns: 0.26 stability
 Previous service: 0.15 stability
 Result names and value patterns showed lower
stability
Highly Variable Features (𝑠(𝑓) < 0.001):
 Length of stay
 Unit stay duration
 Transfer patterns: 0.00002 stability
 Readmission intervals: very low stability
The temporal analysis revealed several key
insights:
 Demographic and administrative features
maintain high stability across encounters but
offer limited discriminative power
 Service-based features provide moderate
stability with better discriminative potential
 Care transition patterns, while less stable,
contain rich information content for temporally
proximate encounters
 Dynamic clinical features show low stability
but high information content, suggesting their
utility for short-term matching
4.3 Hierarchical Feature Weighting
Our hierarchical feature weighting analysis revealed
distinctive patterns in the relationships between
information content, stability, and overall feature
utility for patient matching. Using the three-tier
classification system, features were weighted
according to their stability scores and information
content.
Primary Features (βtier = 1.0) demonstrated
high stability but varied in information content. Four
features qualified for this tier:
 Gender (weight: 997.97, 𝐼

(
𝑓
)
: 1.00 bits)
 Language (weight: 764.93, 𝐼

(
𝑓
)
: 0.76 bits)
 Marital status (weight: 16.36, 𝐼

(
𝑓
)
: 1.60 bits)
 Insurance (weight: 9.67, 𝐼

(
𝑓
)
: 1.48 bits)
Secondary Features (βtier = 0.5) comprised care
unit identifiers. These features showed moderate
stability with information content around 2.5 bits
each:
 First careunit
 Last careunit
Tertiary Features (βtier = 0.25) formed the
largest group with 21 features, including notably:
 Anchor age (highest overall weight: 1526.67,
𝐼

(
𝑓
)
: 6.11 bits)
 Service transition timing (weight: 4.29, 𝐼

(
𝑓
)
:
15.93 bits)
 Transfer patterns (weight: 2.83, 𝐼

(
𝑓
)
: 14.34
bits)
 Readmission intervals (weight: 3.83, 𝐼

(
𝑓
)
:
13.59 bits)
The final weights revealed several important
insights:
 Stability-Information Trade-off: While some
features like service transition timing and
transfer patterns contained high information
content (>14 bits), their lower stability scores
resulted in reduced final weights. Conversely,
demographically stable features like gender
and language achieved higher weights despite
lower information content due to their high
stability.
 Anchor Age Anomaly: Despite being
classified as a tertiary feature, anchor age
achieved the highest overall weight (1526.67)
due to its unique combination of moderate
information content and computational
stability characteristics.
 Care Unit Features: First and last careunit
assignments maintained moderate weights
through balanced stability and information
content, positioning them as reliable secondary
matching criteria.
 Dynamic Features: Highly variable features
like length of stay and unit stay duration,
despite high information content, received
lower weights due to their temporal instability,
aligning with their expected variability across
patient encounters.
This weighting scheme effectively balanced the
trade-off between feature stability and information
content, prioritizing features that maintain consistent
discriminative power across patient encounters while
appropriately discounting unstable or low-
information features.
4.4 Combined Feature Space
Analysis of the feature groups reveals distinct
patterns in discriminative power and information
content. The three major feature groups demonstrate
varying levels of effectiveness for patient matching:
Demographics and Admission Features (𝐷
(
𝐹
)
= 12247.56 bits) that combines the following
features: 'Anchor Age', 'Ggender', 'Language',
ENASE 2025 - 20th International Conference on Evaluation of Novel Approaches to Software Engineering
288

'Anchor Year Group’, ‘Marital Status', and
'Insurance':
 Provides the highest discriminative power
despite relatively low information content
(13.24 bits)
 Dominated by anchor_age and gender
contributions
 Achieves 96% of total discriminative power
across all features
 High stability characteristics enable reliable
long-term matching
ICU Care Patterns (𝐷
(
𝐹
)
= 266.40 bits)
combining 'Los', 'ICU Duration', 'Last Careunit', 'First
Careunit', 'Readmission Intervals', 'Transfer Patterns',
'Service Transition Timing', 'Unit Transitions', 'Unit
Stay Duration':
 Highest information content (94.83 bits)
 Moderate discriminative power driven by
temporal patterns
 Service transitions and readmission intervals
provide key disambiguation capabilities
 Effective for matching temporally proximate
encounters
Clinical Records (𝐷
(
𝐹
)
= 12.10 bits) consisting
of 'Current Service', 'Previous Service', 'Sequential
Measurements', 'Value Patterns', 'Measurement
Patterns', 'Result Name':
 Moderate information content (28.66 bits)
 Limited discriminative power
 Most suitable for secondary verification
 Value in combination with other feature groups
for edge cases
The total discriminative power across all groups
(12526.06 bits) substantially exceeds the theoretical
minimum threshold of log

(
𝑁
)
≈ 16bits for our
population, with demographic features providing the
primary matching power and ICU patterns offering
important secondary discrimination.
Comparison with Existing Approaches
4.5 Comparison with Existing
Approaches
Unlike probabilistic methods, which weight features
empirically (e.g., Ong et al., 2020), our framework
quantifies information content using entropy,
highlighting temporal features’ dominance (e.g.,
15.52 bits for ICU duration vs. 0.998 bits for gender).
Ong et al.’s hybrid approach, combining
deterministic and probabilistic record linkage,
outperformed independent methods by identifying
18%-24% more correct pairs in congenital heart
disease surveillance, leveraging PII like names and
dates. However, their reliance on exact matches
contrasts with our entropy-based feature stability
analysis, which prioritizes temporal patterns over
static identifiers. While Ong et al. addressed data
quality via harmonization, our approach assumes
complete data (e.g., α
(
𝑓
)
=1.0), necessitating future
validation for missingness. Direct accuracy
comparisons remain future work, but Ong’s findings
underscore the potential of hybrid strategies,
suggesting our framework could integrate temporal
stability thresholds to enhance matching robustness
across diverse datasets.
4.6 Limitations and Future Work
Our current approach to temporal anchoring of
demographic features through admission times, while
functional, introduces potential selection bias by
excluding patients without admission records. Future
work should explore alternative temporal reference
points. It should also develop methods to incorporate
outpatient encounters for more comprehensive
patient matching.
A notable limitation is that our feature
combinations often provide discriminative power
significantly exceeding the theoretical minimum of
log

(
𝑁
)
bits - in some cases by several orders of
magnitude. While this redundancy provides
robustness against missing data, it may indicate
computational inefficiency and potential overfitting
to institutional patterns. Future research should
investigate optimal feature selection methods that
balance discriminative power with computational
efficiency while maintaining matching accuracy.
The assumption of demographic stability across
encounters needs careful examination, particularly
for long-term longitudinal studies where
characteristics like insurance status, marital status,
and language preferences may evolve. Research into
dynamic feature weighting that adapts to temporal
distance could enhance matching accuracy.
Missing value patterns significantly impact join
quality between administrative and clinical tables,
potentially skewing stability metrics. Development of
robust imputation methods specifically designed for
temporal healthcare data could address this limitation
and improve feature extraction reliability.
Our stability thresholds, while empirically
derived from MIMIC-IV data, require validation
across diverse healthcare settings and populations.
Multi-institutional studies would help establish
generalizable parameters for feature classification
and weighting.
Information-Theoretic Patient Record Matching in Medical Databases: A Discriminative Power and Feature Analysis Using MIMIC-IV
289

Privacy considerations currently limit cross-
institutional validation of our matching approach.
Future work should incorporate privacy-preserving
computation methods that enable collaborative model
validation without compromising patient
confidentiality.
The current implementation overlooks potential
feature interactions by treating each characteristic
independently. Development of composite features
that capture relationships between administrative,
clinical, and temporal patterns could enhance
discriminative power.
Real-time feature extraction and matching present
computational challenges not addressed in our
retrospective analysis. Research into efficient
algorithms and optimization techniques would
facilitate clinical deployment of our information-
theoretic approach.
Specialty-specific matching requirements and
varying documentation practices across clinical
domains warrant investigation. Adaptive frameworks
that account for department-specific feature stability
and information content could improve matching
accuracy in specialized care settings.
Future work could explore hybridizing our
entropy-based framework with deterministic
methods, as Ong et al. suggest, to address missing
data and validate across diverse healthcare systems.
5 CONCLUSIONS
The information-theoretic analysis successfully
established a framework for patient matching in
critical care settings, revealing three complementary
feature groups: demographics/administrative
( 𝐷
(
𝐹
)
=12247.56 bits), ICU care patterns
( 𝐷
(
𝐹
)
=266.40 bits), and clinical records
( 𝐷
(
𝐹
)
=12.10 bits). While the combined
discriminative power (12526.06 bits) substantially
exceeds the theoretical minimum threshold
( log

(
𝑁
)
≈16bits), this significant redundancy
presents both advantages and challenges.
The excess discriminative power provides
robustness against missing data and institutional
variability. However, it suggests potential
computational inefficiencies and possible overfitting
to institution-specific patterns. Future
implementations should focus on optimizing feature
selection to maintain matching accuracy while
reducing computational overhead.
The research demonstrates that effective patient
matching requires balancing:
 Feature stability vs. information content
 Computational efficiency vs. redundancy
 Institutional generalizability vs. local pattern
optimization
This framework provides a foundation for
implementing reliable patient matching systems,
though further validation across diverse healthcare
environments and optimization of feature selection
methods is needed.
REFERENCES
Blecker, S., Katz, S. D., Horwitz, L. I., Kuperman, G., Park,
H., Gold, A., & Sontag, D. (2016). Comparison of
Approaches for Heart Failure Case Identification From
Electronic Health Record Data. JAMA Cardiology,
1(9), 1014. https://doi.org/10.1001/
jamacardio.2016.3236
Clark, S. J., Halter, M., Porter, A., Smith, H. C., Brand, M.,
Fothergill, R., Lindridge, J., McTigue, M., & Snooks,
H. (2019). Using deterministic record linkage to link
ambulance and emergency department data: Is it
possible without patient identifiers?: A case study from
the UK. International Journal of Population Data
Science, 4(1). https://doi.org/10.23889/ijpds.v4i1.1104
Duggal, R., Khatri, S. K., & Shukla, B. (2015). Improving
patient matching: Single patient view for Clinical
Decision Support using Big Data analytics. 2015 4th
International Conference on Reliability, Infocom
Technologies and Optimization (ICRITO) (Trends and
Future Directions), 1–6. https://doi.org/10.1109/
ICRITO.2015.7359269
Evans, R. S., Benuzillo, J., Horne, B. D., Lloyd, J. F.,
Bradshaw, A., Budge, D., Rasmusson, K. D., Roberts,
C., Buckway, J., Geer, N., Garrett, T., & Lappé, D. L.
(2016). Automated identification and predictive tools to
help identify high-risk heart failure patients: Pilot
evaluation. Journal of the American Medical
Informatics Association, 23(5), 872–878.
https://doi.org/10.1093/jamia/ocv197
Fernandes, L., & O’Connor, M. (2015). Accurate Patient
Identification—A Global Challenge. Perspectives in
Health Information Management, International, 1–7.
Gehrmann, S., Dernoncourt, F., Li, Y., Carlson, E. T., Wu,
J. T., Welt, J., Foote, J., Moseley, E. T., Grant, D. W.,
Tyler, P. D., & Celi, L. A. (2018). Comparing deep
learning and concept extraction based methods for
patient phenotyping from clinical narratives. PLOS
ONE, 13(2), e0192360. https://doi.org/10.1371/
journal.pone.0192360
Godlove, T., & Ball, A. W. (2015). Patient Matching within
a Health Information Exchange. https://pmc.ncbi.nlm.
nih.gov/articles/PMC4696093/
Goldberger, A. L., Amaral, L. A., Glass, L., Hausdorff, J.
M., Ivanov, P. C., Mark, R. G., Mietus, J. E., Moody,
G. B., Peng, C.-K., & Stanley, H. E. (2000).
PhysioBank, PhysioToolkit, and PhysioNet:
ENASE 2025 - 20th International Conference on Evaluation of Novel Approaches to Software Engineering
290

Components of a new research resource for complex
physiologic signals. Circulation, 101(23), e215–e220.
Hua, Y., Wang, L., Nguyen, V., Rieu-Werden, M.,
McDowell, A., Bates, D. W., Foer, D., & Zhou, L.
(2023). A Deep Learning Approach for Transgender
and Gender Diverse Patient Identification in Electronic
Health Records. Journal of Biomedical Informatics,
147, 104507.
Johnson, A., Bulgarelli, L., Pollard, T., Gow, B., Moody,
B., Horng, S., Celi, L. A., & Mark, R. (2024). MIMIC-
IV (Version 3.1) [Dataset]. PhysioNet.
https://doi.org/10.13026/KPB9-MT58
Johnson, A. E. W., Bulgarelli, L., Shen, L., Gayles, A.,
Shammout, A., Horng, S., Pollard, T. J., Hao, S.,
Moody, B., Gow, B., Lehman, L. H., Celi, L. A., &
Mark, R. G. (2023). MIMIC-IV, a freely accessible
electronic health record dataset. Scientific Data, 10(1),
1. https://doi.org/10.1038/s41597-022-01899-x
Just, B. H., Marc, D., Munns, M., & Sandefer, R. (2016).
Why patient matching is a challenge: Research on
master patient index (MPI) data discrepancies in key
identifying fields. Perspectives in Health Information
Management, 13(Spring).
Kho, A. N., Cashy, J. P., Jackson, K. L., Pah, A. R., Goel,
S., Boehnke, J., Humphries, J. E., Kominers, S. D.,
Hota, B. N., Sims, S. A., Malin, B. A., French, D. D.,
Walunas, T. L., Meltzer, D. O., Kaleba, E. O., Jones, R.
C., & Galanter, W. L. (2015). Design and
implementation of a privacy preserving electronic
health record linkage tool in Chicago. Journal of the
American Medical Informatics Association, 22(5),
1072–1080. https://doi.org/10.1093/jamia/ocv038
McCoy, A. B., Wright, A., Kahn, M. G., Shapiro, J. S.,
Bernstam, E. V., & Sittig, D. F. (2013). Matching
identifiers in electronic health records: Implications for
duplicate records and patient safety. BMJ Quality &
Safety, 22(3), 219–224. https://doi.org/10.1136/bmjqs-
2012-001419
Nie, D., & Roantree, M. (2019). Detecting Multi-
Relationship Links in Sparse Datasets: Proceedings of
the 21st International Conference on Enterprise
Information Systems, 149–157. https://doi.org/
10.5220/0007696901490157
Ong, T. C., Duca, L. M., Kahn, M. G., & Crume, T. L.
(2020). A hybrid approach to record linkage using a
combination of deterministic and probabilistic
methodology. Journal of the American Medical
Informatics Association, 27(4), 505–513.
https://doi.org/10.1093/jamia/ocz232
Riplinger, L., Piera-Jiménez, J., & Dooling, J. P. (2020).
Patient Identification Techniques – Approaches,
Implications, and Findings. Yearbook of Medical
Informatics, 29(01), 081–086. https://doi.org/10.1055/
s-0040-1701984
Sehili, Z., Kolb, L., Borgs, C., Schnell, R., & Rahm, E.
(2015). Privacy preserving record linkage with PPJoin.
Shannon, C. E. (1948). A mathematical theory of
communication. The Bell System Technical Journal,
27(4), 623–656. https://doi.org/10.1002/j.1538-
7305.1948.tb00917.x
Toth, C., Durham, E., Kantarcioglu, M., Xue, Y., & Malin,
B. (2014). SOEMPI: a secure open enterprise master
patient index software toolkit for private record linkage.
AMIA Annual Symposium Proceedings, 2014, 1105.
Vatsalan, D., Christen, P., & Rahm, E. (2017). Scalable
Multi-Database Privacy-Preserving Record Linkage
using Counting Bloom Filters (No. arXiv:1701.01232).
arXiv. https://doi.org/10.48550/arXiv.1701.01232
Zech, J., Husk, G., Moore, T., & Shapiro, J. (2016).
Measuring the Degree of Unmatched Patient Records in
a Health Information Exchange Using Exact Matching.
Applied Clinical Informatics, 07(02), 330–340.
https://doi.org/10.4338/ACI-2015-11-RA-0158
Information-Theoretic Patient Record Matching in Medical Databases: A Discriminative Power and Feature Analysis Using MIMIC-IV
291