Knowledge-Enriched Structured EHR Features for 30-Day Hospital Readmission Prediction on MIMIC-IV
Mohamad Najafi, Hongyun Fu, Mathias Brochhausen, Jian Wu, Yaohang Li
cs.LG · q-bio.QM
Quick verdict
Knowledge-enriched structured EHR features hit AUROC 0.743 (CV) on MIMIC-IV readmission, competitive with note-based models; demographics dominate the gain.
Summary
This paper proposes a sparse, knowledge-enriched representation of structured EHR data for predicting 30-day hospital readmission without clinical notes. From MIMIC-IV v2.2 the authors build a cohort of 422,622 admissions (176,901 patients; 20.1% readmission rate), mapping diagnoses to MONDO ontology ancestor levels (L1–L3), procedures to AHRQ CCS categories, prescriptions to RxNorm ingredient indicators, and lab events to organ-system aggregates, yielding a 23,430-dimensional matrix that is 99.78% sparse. This configuration (B2) is compared against BioClinicalBERT-encoded features (B1) and a combined variant (B3) across six classifiers under a patient-level 60/20/20 split. LightGBM on B2 performs best: AUROC 0.7412 (AUPRC 0.441) on the held-out test set and 0.743±0.007 under 20-fold cross-validation, exceeding B1 by +0.028 AUROC (bootstrap 95% CI [+0.026, +0.031]); however, linear and neural models favor B1, and B3 adds no consistent benefit. Ablations show demographics dominate (48.5% of model gain; 0.6985 AUROC alone), while deeper MONDO levels (L2/L3) contribute negligibly. Main limitations are a single-center cohort, a cross-study (not controlled) comparison with note-based results on a differently filtered cohort, cross-validation for the best configuration only, and a labeling convention that assigns every patient's last admission a negative label.
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 knowledge-enriched sparse structured representation (B2, 23,430 dims, 99.78% sparse) achieves AUROC 0.7412 on the held-out test set and 0.743±0.007 under patient-level 20-fold CV for 30-day readmission on MIMIC-IV, competitive with note-based methods without using notes.
- Ontology-enriched features outperform BioClinicalBERT-encoded features for gradient-boosted trees (LightGBM +0.028 AUROC, bootstrap 95% CI [+0.026, +0.031], p<0.001), while linear and neural models prefer the BERT features and the combined B3 configuration adds no consistent benefit.
- Each feature dimension corresponds to a named clinical concept, giving interpretability at roughly 24× less storage (176 MB vs 4.2 GB) and no language-model inference cost.
- Interpretability analysis shows demographics dominate (48.5% of LightGBM gain; 0.6985 AUROC alone), organ-system labs, RxNorm ingredients, unmapped ICD codes, and MONDO L1 categories contribute meaningfully, and deeper MONDO levels (L2/L3) contribute negligibly.
Strengths
Leakage-controlled large-cohort design
§4: 422,622 admissions / 176,901 patients with a patient-level, label-stratified 60/20/20 split (253,254/84,328/85,040); unlike the key comparator, the cohort is not conditioned on discharge-summary availability, so the deployment scenario (no notes) is honest.
Controlled three-way feature comparison
§3/§6: B1 (BioClinicalBERT), B2 (ontology), B3 (combined) share identical cohort, label, and split across six classifiers; dimension accounting checks out (93+527+24+1,324+21,462 = 23,430), and bootstrap deltas are reported for the key contrasts (§7).
Multi-method interpretability with a consistent story
Subgroup audit with confidence intervals
Quantified practical efficiency
Weaknesses
Demographics dominate, tempering the knowledge-enrichment headline
The largest block is not knowledge-derived
Selective uncertainty reporting
Cross-study comparability is uncontrolled
B3 hypothesis unsupported and ablation representation mismatch
Full Review Report
$1.99 · one-timeUnlock the complete preliminary analysis Evolve produced for this paper.
Report preview
Scope & Claims
The paper targets 30-day all-cause hospital readmission prediction on MIMIC-IV v2.2 using structured EHR data only — no clinical notes — and asks whether explicit medical-knowledge features can match transformer-encoded representations. Central claims: (1) a 23,430-dimensional sparse matrix (B2) built by mapping diagnoses to MONDO ancestor levels L1–L3 (6,557 features) plus 14,905 unmapped-ICD one-hots, procedures to 527 AHRQ CCS categories, prescriptions to 1,324 RxNorm ingredient indicators, and lab events to 24 organ-system aggregates; (2) LightGBM on B2 reaches AUROC 0.7412 on the held-out 20% test split (AUPRC 0.441) and 0.743±0.007 under patient-level 20-fold CV on a 422,622-admission cohort (20.1% prevalence); (3) this is "competitive with, not superior to" note-based results such as Almeida et al.'s 0.727; and (4) demographics dominate prediction while deep ontology levels contribute nothing.
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 AUROC does this paper achieve for 30-day readmission prediction on MIMIC-IV?
The best configuration — LightGBM on the knowledge-enriched structured features (B2) — reaches AUROC 0.7412 on a held-out 20% test split (AUPRC 0.441) and 0.743±0.007 under 20-fold cross-validation, on a cohort of 422,622 admissions with a 20.1% readmission rate. The authors position this as comparable to note-based methods (e.g., 0.727 by Almeida et al. on a different, note-conditioned cohort), explicitly not superior.
Does the model use clinical notes or NLP?
No. It uses only structured EHR data — ICD diagnosis and procedure codes, lab events, prescriptions, and demographics — mapped through four knowledge sources: the MONDO disease ontology (three ancestor levels), the AHRQ CCS procedure classification, RxNorm ingredient-level drug indicators, and keyword-based organ-system lab aggregation. This avoids dependence on discharge summaries and language-model inference costs.
Which features matter most for the readmission prediction?
Per the paper's ablation (Table 3), administrative/demographic features dominate: 48.5% of model gain, and 0.6985 AUROC on their own versus 0.7412 for the full model. Organ-system labs, drug ingredients, and unmapped ICD codes each add 1–2.3 permutation points, and only first-level MONDO disease categories (L1) help among ontology features; deeper levels (L2/L3) contribute essentially nothing.
Is code or data available for this paper?
The paper uses the credentialed public MIMIC-IV v2.2 dataset (PhysioNet, requires a data use agreement) and standard tools (LightGBM, XGBoost, FAISS), but the text states no code repository or release. Ontology-mapping heuristics (ICD truncation, RxNorm relationship traversal, lab keyword matching) are described only in prose, so exact reimplementation requires some reconstruction.
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 1–3 are extracted as fragmented text (captions and axis labels); Figure 3's numeric labels are readable but figure rendering cannot be verified.
- The abstract in the extracted text is truncated mid-sentence; conclusions are drawn from the full body sections, which appear complete including references.
- Cross-study baseline numbers (Almeida et al. 0.704/0.727, and other cited AUROCs) are taken from the paper's own citations and were not independently verified here.
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.