Appendix: Content Summaries of Selected Best Papers for the IMIA Yearbook 2023 Section
“Clinical Information Systems”[1]
Bialke M, Geidel L, Hampf C, Blumentritt A, Penndorf P, Schuldt R, Moser FM, Lang
S, Werner P, Stäubert S, Hund H, Albashiti F, Gührer J, Prokosch HU, Bahls T, Hoffmann
W
A FHIR has been lit on gICS: facilitating the standardised exchange of informed consent
in a large network of university medicine
BMC Med Inform Decis Mak 2022 Dec 19;22(1):335. doi: 10.1186/s12911-022-02081-4
Overview:
The article discusses the technical implementation of patient consent management in
the context of the NUM-CODEX project, which aims to enable cross-site data exchange
via FHIR in the area of consent. The authors describe the use of globally unique object
identifiers (OID) and semantic statements to develop a common representation of the
MII Broad Consent, which was extended with NUM-specific extensions. They also highlight
the challenges of implementing complex FHIR consent profiles and the need for extensive
practical experience and quality control to ensure syntactically and semantically
correct implementation.
Detailed Description:
Goal
The goal of the text is to describe the implementation of a FHIR-compliant and interoperable
nationwide exchange of consent information using gICS and TTP-FHIR Gateway.
Methods
The solution covers requirements identified in NUM-CODEX setting, which is part of
a network of university medicines (NUM) to support COVID-19 and pandemic research
at national level.
All 34 participating university hospitals work upon harmonized infrastructural as
well as legal basis for their data protection compliant collection and transfer.
Informed consent from patients is required for processing their health data at NUM
sites, transfer to CODEX, use & access procedure based on GDPR Art. 6(1) lit. a).
Findings
A reliable consent management system was successfully implemented by University Medicine
Greifswald using MII broad consent in mid2020.
Current developments in the FHIR community must be considered when implementing cross-site
data exchange via FHIR consents.
Steps were taken towards conception/implementation standardized solutions for provisioning
consents through gICS with TTP-FHIT gateway serving as an intermediary between external
infrastructure components.
Implications
Technical prerequisites have been achieved with help from gICS/TTP-FHIT Gateway enabling
FHIT compliant provision info about patient's informed consents across multiple sites
while maintaining compliance w/GDPR regulations.
Customised templates simplify assurance technical interoperability among all NUM/MII
sites.
Guardiolle V, Bazoge A, Morin E, Daille B, Toublant D, Bouzillé G, Merel Y, Pierre-Jean
M, Filiot A, Cuggia M, Wargny M, Lamer A, Gourraud PA
Linking Biomedical Data Warehouse Records With the National Mortality Database in
France: Large-scale Matching Algorithm
JMIR Med Inform. 2022 Nov 1;10(11):e36711. doi: 10.2196/36711
Overview:
The paper describes a large-scale matching algorithm that links biomedical data warehouse
(BDW) records with the French National Mortality Database (FNMD) to determine vital
status after discharge, which is crucial for medical research. The algorithm uses
advanced data cleaning and knowledge of the naming system, along with the Damerau-Levenshtein
distance (DLD), to overcome challenges such as absence of unique common identifiers,
name changes, and clerical errors. The algorithm's performance was evaluated using
BDW data from three university hospitals, and the results showed that the DLD-based
algorithm outperformed the direct algorithm in terms of sensitivity/recall by 11%.
Detailed Description:
Goal
The goal of the text is to present a study that developed and evaluated two algorithms
for linking patient records between local biomedical data warehouses (BDWs) and the
French National Mortality Database (FNMD).
Methods
The study utilized record linkage techniques, advanced data cleaning, knowledge of
naming systems, Damerau-Levenshtein distance calculations, and blocking techniques.
Two algorithms were developed: a direct-matching algorithm and a deterministic matching
algorithm based on DLD.
Findings
Matching large-scale BDW records with FNMD is challenging due to absence of unique
common identifiers between databases.
The use of an advanced deterministic matching algorithm such as the DLD-based algorithm
showed higher sensitivity/recall than direct matching algorithms.
Specificity was ≥98%, reducing risk differential biases between groups.
The execution time per patient decreased as total number increased but blocking technique
reduced required comparisons by at least 40k times.
Implications
The findings have implications for medical research using open-source external data
sources like FNMD in improving usage value BDWs.
The proposed method can be used routinely update vital status in BDW records from
FNMD on large scale
Deterministic methods are more efficient when dealing with datasets without unique
common identifiers
Poelzl G, Egelseer-Bruendl T, Pfeifer B, Modre-Osprian R, Welte S, Fetz B, Krestan
S, Haselwanter B, Zaruba MM, Doerler J, Rissbacher C, Ammenwerth E, Bauer A
Feasibility and effectiveness of a multidimensional post-discharge disease management
programme for heart failure patients in clinical practice: the HerzMobil Tirol programme
Clin Res Cardiol 2022 Mar;111(3):294-307. doi: 10.1007/s00392-021-01912-0
Overview:
The HerzMobil Tirol program is a multidimensional post-discharge disease management
program for heart failure patients that incorporates a telemedical monitoring system
and a comprehensive network of specialized heart failure nurses, resident physicians,
and referral centers. In a non-randomized study of 508 acute heart failure patients,
the program was found to be feasible and effective in reducing the primary endpoint
of time to HF readmission and all-cause mortality within 6 months by 46% compared
to usual care. The program also showed a reduction in the composite of recurrent HF
hospitalization and death within 6 months and a lower mortality rate after 1 year.
Detailed Description:
Goal
To evaluate the feasibility and effectiveness of a multidimensional post-discharge
disease management program for heart failure patients using telemedical technology.
Methods
The study included 508 acute heart failure (AHF) patients managed by HerzMobil Tirol
or usual care after discharge from hospital.
Patients in the HMT group underwent standardized evaluation, received structured follow-up,
and were provided with a blood pressure and heart rate monitor, weighing scale, and
smartphone for daily data acquisition.
Logistic regression models were used to calculate hazard ratios for primary outcomes
including 6-month HF hospitalization and all-cause mortality at 6 months.
Findings
Management by HerzMobil Tirol was associated with a significant reduction (46%) in
time to HF readmission within six months compared with usual care.
The composite endpoint was significantly lower with HMT than UC indicating that the
program is effective.
Patients compliance among participants remaining in the HMT programme was high; only
six out of two hundred fifty-one patients were found negligent on data transfer but
remained until completion after three months.
Implications
Results suggest that specific disease management programs should be implemented widely
following an acute heart failure event as they can prevent readmissions while improving
clinical outcomes such as self-care behavior among patients.
Zou Y, Pesaranghader A, Song Z, Verma A, Buckeridge DL, Li Y
Modeling electronic health record data using an end-to-end knowledge-graph-informed
topic model
Sci Rep 2022 Oct 25;12(1):17868. doi: 10.1038/s41598-022-22956-w
Overview:
The article describes a novel approach for modeling electronic health record (EHR)
data using an end-to-end knowledge-graph-informed topic model called Graph ATtention-Embedded
Topic Model (GAT-ETM). The GAT-ETM method leverages a medical knowledge graph to learn
clinically meaningful disease topics from EHR data, and demonstrates superior performance
over alternative methods for tasks such as drug imputation and disease diagnosis prediction.
The approach has potential applications in computational phenotyping, patient stratification,
and drug recommendations, and provides a promising avenue for refining disease phenotypes
and discovering novel disease comorbidities from large-scale EHR datasets.
Detailed Description:
Goal
The goal of the text is to present Graph ATtention-Embedded Topic Model (GAT-ETM),
an end-to-end taxonomy-knowledge-graph-based multimodal embedded topic model that
distills latent disease topics from EHR data by learning the embedding from a constructed
medical knowledge graph.
Methods
GAT-ETM assumes a generative process for each patient in the EHR corpus, where topic
mixture membership is drawn from logistic-normal and categorical distributions.
To extract meaningful and interpretable disease topics, a linear decoder is used to
reconstruct EHR data such that the linear projections can directly map to individual
latent topics.
A graph augmentation strategy was proposed by connecting nodes with their ancestry
nodes along taxonomy in order to maximize information flow among EHR nodes on the
graph.
An ICD-ATC knowledge graph is leveraged to learn code embedding.
Empirically, it was found that mini-batch stochastic gradient descent worked well
for updating models with large datasets.
Findings
GAT-ETM demonstrated superior performance over alternative methods on all tasks including
drug imputation and disease diagnosis prediction.
GAT-ETM learned clinically meaningful embeddings of codes and discovered accurate
patient representations for stratification purposes as well as drug recommendations.
The study evaluated model performance using expert-derived rule-based labels across
12 chronic diseases; results showed high accuracy classification scores across all
diseases tested.
Implications
Incorporating multiple views through integrating knowledge graphs brings complementary
information which improves phenotyping accuracy quantitatively & qualitatively.
Attention mechanism enables tracking input feature contributions which will be useful
in understanding connections between different diseases.
