In Medical Claims Data, Enhancing Predictive Performance for Major Adverse Cardiovascular Events Using Cross Attention
Yuhei Fujioka, Daitaro Misawa, Tatsuyoshi Ikenoue, Shingo Fukuma
cs.LG
Quick verdict
Proposes cross-attention over diagnosis-treatment links; MACE ROC-AUC 0.7720 beats self-attention 0.7586, but evidence rests on one private Japanese cohort.
Summary
The paper proposes a deep learning model for predicting major adverse cardiovascular events (MACE) from Japanese health insurance data, combining eight tabular health-checkup features with twelve months of medical claims (diagnoses, procedures, prescriptions). Its central idea is a cross-attention block that treats diagnoses as queries and treatments as keys/values, learning many-to-many diagnosis–treatment relationships that prior self-attention approaches (e.g., ExBEHRT) force into many-to-one links. Medical codes are decomposed into hierarchical sub-tokens (ICD-10 levels, procedure and prescription categories) and processed per month by transformer encoders. On 51,367 subjects from a single Japanese insurer (2.38% MACE incidence), with 9-fold cross-validation, the proposed model achieves ROC-AUC 0.7720 (MCC 0.1525) versus 0.7586 for a self-attention variant, 0.7559 for LightGBM with claims features, and 0.7521 for the applied Pooled Cohort Equations. Wilcoxon (p=0.0039) and DeLong (p=0.0159, ensemble 0.7791 vs 0.7653) tests support the gap. Acknowledged limitations include removal of duplicate codes, no use of temporal ordering, exclusion of code-heavy patients, and a predominantly male construction-industry cohort; code and data are not released.
Review complete
Assessment at a glance
5
risk flags identified
● high 0
● medium 3
● low 2
5 strengths · 5 weaknesses · 5 author questions
AI text signal = Binoculars zero-shot detector score over 20 sampled passages (Fast-DetectGPT co-reported) — absence does not prove human authorship, and a flag does not imply misconduct.
Levels reflect verification complexity — not a quality score. Evolve never rates your paper.
Key claims under review
- A cross-attention block with diagnoses as queries and treatments as keys/values learns many-to-many diagnosis–treatment relationships from claims data that self-attention models force into many-to-one links (§1.4, §3.1.1).
- The proposed model reaches ROC-AUC 0.7720 and MCC 0.1525 (mean over 9-fold CV) on 51,367 Japanese insured subjects, exceeding its self-attention variant (0.7586/0.1253), LGBM with claims features (0.7559), and the applied Pooled Cohort Equations (0.7521) (Table 3).
- Adding medical claims data to health-checkup-only inputs improves MACE prediction for every learned model tested (Table 3: FFN 0.7412 vs SA 0.7586 vs LGBM 0.7374→0.7559).
- The CA-over-SA gain is statistically significant by fold-level Wilcoxon testing (p=0.0039 for both metrics) and by DeLong test on ensembled predictions (p=0.0159; ensemble AUC 0.7791 vs 0.7653) (§5.1, Appendix H).
Strengths
Well-motivated architectural contribution
The many-to-many diagnosis–treatment argument (§1.1, §2.3) directly targets a real limitation of ExBEHRT's many-to-one linking; cross-attention with diagnoses as queries and treatments as key/value (Figure 2, §3.1.1) is a clean fit for it.
Clean ablation ladder
Table 3 isolates contributions stepwise: FFN on checkup data only (0.7412) → +claims via self-attention (0.7586) → +claims via cross-attention (0.7720), so the data-addition and mechanism effects are separately observable.
Unusually transparent appendix documentation
Honest baseline disclosure
Reusable preprocessing scheme
Weaknesses
Fragile statistical grounding
Clinically thin evaluation
Attribution not fully isolated
No interpretability evidence
Severely limited external validity and reproducibility
Full Review Report
$1.99 · one-timeUnlock the complete preliminary analysis Evolve produced for this paper.
Report preview
Scope & Claims
This AIDSH-KDD '24 workshop paper targets prediction of major adverse cardiovascular events — hospitalization or death from AMI (ICD-10 I20–I25), cerebrovascular disease (I60–I69), heart failure (I50), or peripheral arterial disease (I70); Table 1 — using FY2014 health checkup data plus one year of medical claims. The central claim is architectural: because claims data do not explicitly link diagnoses to treatments and the true relationship is many-to-many (e.g., sacubitril/valsartan serving two diagnoses), a cross-attention block — diagnoses as queries, treatments as keys and values (Figure 2) — should outperform self-attention models that must treat the relationship as many-to-one (the ExBEHRT strategy, §2.3). Secondary claims: adding claims data to checkup data helps every learned model, and the CA-over-SA gain is significant (Wilcoxon p=0.0039; DeLong p=0.0159 on ensembles, Appendix H). Claims are stated conservatively and the Limitations section (§7, Appendix J) is unusually candid.
The remaining sections cover Methodology Assessment, Statistical Evidence, Results & Reproducibility, Limitations & Risks, and Verdict & Recommendations.
- All 5 strengths & 5 weaknesses, analyzed in detail
- 5 risk flags with severity, section & reviewer questions
- Methods & statistical evidence assessment
- 5 questions a peer reviewer would ask
- Full six-dimension evaluation report & recommendations
One-time purchase · this paper only
One-time purchase, this paper only. AI-generated preliminary analysis — not peer review.
Frequently asked questions
What is the main contribution of this paper?
It applies a cross-attention mechanism to Japanese medical claims data so that diagnoses and treatments mutually inform each other's representations (many-to-many relationships), improving prediction of major adverse cardiovascular events (MACE) over self-attention models that link each treatment to only one diagnosis.
What performance did the model achieve?
ROC-AUC 0.7720 and MCC 0.1525 (mean over 9-fold cross-validation) on 51,367 individuals with 2.38% MACE incidence, versus 0.7586 for its self-attention variant, 0.7559 for LightGBM with claims features, and 0.7521 for the applied Pooled Cohort Equations.
Is the dataset or code publicly available?
No. The data come from a single Japanese health insurance association and the paper contains no code or data availability statement, so the results cannot currently be independently reproduced.
Can this model be used in clinical practice today?
Not directly. It was trained and evaluated on one occupational insurer's cohort (about 73% male construction-industry workers), has no external validation, no calibration or threshold analysis, and excludes patients with many medical codes — all of which would need addressing before any clinical use.
About this review
This is an AI-generated preliminary review by Evolve.Pub, produced from the openly available text of the paper. It is not peer review and does not represent the authors' views. The canonical record of the paper is at the publisher link above.
- Figures 1, 2, A.3, A.4 and G.5 are present only as captions; their graphical content (architecture detail, cross-attention flow, and the exclusion counts in the selection flow chart) cannot be verified from text extraction.
- The set-builder notation in §3.2.2 is garbled by PDF extraction, but the surrounding prose makes the intended multi-list construction interpretable.
Go deeper on this paper
Wrote a paper? Stress-test it before reviewers do.
Upload your manuscript for the same structured scrutiny — reasoning, statistics, evidence, figures, and reproducibility.