Research map / Label-efficient and robust learning / Recognition and long-tail
NeuralCohort: Cohort-aware Neural Representation Learning for Healthcare Analytics
Read the paperCheck your idea against 108,967 papers
AI-extracted summary. Written by AI from the paper's full text. Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.
Proposes NeuralCohort, a plug-in cohort-aware representation learner that derives fine-grained patient cohorts from pseudo patient similarity and fuses intra- and inter-cohort graph information into EHR backbone representations.
Problem
Existing EHR representation learning and cohort studies struggle to derive fine-grained patient cohorts and jointly exploit intra-cohort and inter-cohort information for healthcare prediction tasks.
Approach
Pre-context Cohort Synthesis learns patient representations with a pseudo similarity (weighted Jaccard over diagnosis codes with time decay) via mutual-information contrastive training, then derives cohorts by soft clustering; Biscale Cohort Learning fuses a local intra-cohort GCN and a global inter-cohort encoder-decoder with backbone representations through cross-domain attention.
Key results
- NeuralCohort improves ClinicalBERT on MIMIC-III readmission from 0.651 to 0.681 AUROC.
- Up to 8.1% AUROC gain across backbones, with up to 8.0% AUPRC and 16.3% accuracy gains reported.
- Calinski-Harabasz score for Med2Vec rises from 18.4 to 25.7 with NeuralCohort on long LOS prediction.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Integrate NeuralCohort | ClinicalBERT · MIMIC-III readmission | AUROC: 0.651±0.006 -> 0.681±0.005 | Helped |
| Integrate NeuralCohort | Med2Vec · MIMIC-III readmission | AUROC: 0.614±0.004 -> 0.634±0.005 | Helped |
| Integrate NeuralCohort | MiME · MIMIC-III readmission | AUROC: 0.602±0.005 -> 0.629±0.003 | Helped |
| Integrate NeuralCohort | ClinicalBERT · MIMIC-III long LOS | AUROC: 0.590±0.002 -> 0.671±0.004 | Helped |
| Integrate NeuralCohort | Med2Vec · MIMIC-III long LOS | AUROC: 0.894±0.001 -> 0.906±0.004 | Helped |
| Integrate NeuralCohort | MiME · MIMIC-III long LOS | AUROC: 0.904±0.003 -> 0.923±0.002 | Helped |
| Add Pre-context Cohort Synthesis Module (ablation) | NeuralCohort with ClinicalBERT · MIMIC-III long LOS | AUPRC: 0.682 -> 0.738 | Helped |
| Add Local Intra-cohort Modeling (ablation) | NeuralCohort with ClinicalBERT · MIMIC-III long LOS | AUPRC: 0.675 -> 0.738 | Helped |
| Add Global Inter-cohort Modeling (ablation) | NeuralCohort with ClinicalBERT · MIMIC-III long LOS | AUPRC: 0.714 -> 0.738 | Helped |
Surprising findings
- Traditional cohort construction (gender, age, diabetes, hypertension) often performs comparably to or worse than the backbone alone.
- Several baselines (e.g., DGLoS, DEC, DEKM) degrade AUROC on some backbone/task combinations.
Limitations
- Cohort derivation method is stated to be outside the paper's primary scope, so clustering choice is not optimized.
- Evaluated only on EHR datasets and two prediction tasks (readmission and long LOS).
Experimental setup
Datasets: MIMIC-III MIMIC-IV Diabetes130
Benchmarks: MIMIC-III readmission prediction MIMIC-III long LOS prediction MIMIC-IV Diabetes130
Metrics: AUPRC AUROC accuracy Calinski-Harabasz score
Baselines: KNN K-Means DEC DEKM GRASP DGLoS IDC
Builds on: Med2Vec MiME ClinicalBERT GRASP DEC
Closest papers by meaning
- ACES: Automatic Cohort Extraction System for Event-Stream Datasets (ICLR 2025 · similarity 0.88)
- Exploiting Negative Samples: A Catalyst for Cohort Discovery in Healthcare Analytics (ICML 2024 · similarity 0.85)
- Context Clues: Evaluating Long Context Models for Clinical Prediction Tasks on EHR Data (ICLR 2025 · similarity 0.83)
- DearLLM: Enhancing Personalized Healthcare via Large Language Models-Deduced Feature Correlations (AAAI 2025 · similarity 0.83)
- Contrastive Learning for Clinical Outcome Prediction with Partial Data Sources (ICML 2024 · similarity 0.82)
- Integrating Sequence and Image Modeling in Irregular Medical Time Series Through Self-Supervised Learning (AAAI 2025 · similarity 0.82)
Writing a paper in this area? scime draft is a free LaTeX editor with AI included, real-time co-authors and a check of your references and venue rules before you submit. It also imports OpenAI Prism projects.