Research map / Label-efficient and robust learning / Recognition and long-tail
Exploiting Negative Samples: A Catalyst for Cohort Discovery in 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 discovering patient cohorts among negative samples by computing data Shapley values, building a Negative Sample Shapley Field, transforming it with an SDAE carrying an isotropy constraint, and clustering with DBSCAN, applied to hospital-acquired AKI prediction on NUH EMR data.
Problem
Negative samples in binary healthcare diagnosis/prognosis tasks are open-ended and underexplored, yet they may contain patients with related conditions (comorbidities, complications) that merit clinical attention; existing cohort methods focus on positives.
Approach
Treat each negative sample's data Shapley value as a scalar field over negative samples, learn a manifold representation with a structure-preserving SDAE plus an isotropy penalty on Shapley value changes, and run DBSCAN on high-value samples to find cohorts.
Key results
- Using only negative samples with positive data Shapley values improved AUC for all five classifiers (LR, GBDT, AdaBoost, RF, MLP) versus using all negatives.
- SDAE manifold learning reduced 709 lab-test dimensions to 64 and lowered AUC by roughly 5% for most classifiers relative to raw inputs.
- DBSCAN on the top 40% high-Shapley negatives yielded seven clinically interpreted cohorts (e.g., inflammatory, hepatic/hematological, respiratory failure/metabolic acidosis), validated by clinicians.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Exclude negative samples with data Shapley values below 0 | AKI prediction classifiers · Hospital-acquired AKI | AUC | Helped |
| Use SDAE-decoded representations instead of raw input features (ablation) | AKI prediction classifiers · Hospital-acquired AKI | AUC: approximately 5% | Hurt |
| Use influence-function values instead of data Shapley values to select negative samples | AKI prediction · Hospital-acquired AKI | AUC | Hurt |
| Add an isotropy constraint to manifold learning (ablation) | Negative Sample Shapley Field cohort discovery · Hospital-acquired AKI | Cohort discovery | Helped |
Surprising findings
- Negative samples with negative data Shapley values are likely to be future positives rather than mislabeled data, per the authors' interpretation for AKI.
Limitations
- Evaluation on a single hospital's EMR data for one disease (AKI), with clinical validation by clinicians rather than quantitative outcomes.
- Data Shapley computation requires retraining classifiers many times, which is computationally costly.
Experimental setup
Datasets: National University Hospital (Singapore) EMR data MIMIC-III (supplementary)
Benchmarks: NUH hospital-acquired AKI prediction dataset MIMIC-III (supplementary cohort analysis)
Metrics: AUC
Baselines: influence function-based data valuation contrastive PCA DCN DEKM PU learning (Elkan-Noto, bagging PU)
Builds on: data Shapley value stacked denoising autoencoder DBSCAN influence functions
Closest papers by meaning
- NeuralCohort: Cohort-aware Neural Representation Learning for Healthcare Analytics (ICML 2025 · similarity 0.85)
- Risk Horizons: Structured Hypothesis Spaces for Longitudinal Clinical Prediction (NeurIPS 2026 · similarity 0.84)
- ACES: Automatic Cohort Extraction System for Event-Stream Datasets (ICLR 2025 · similarity 0.84)
- Breaking the Barrier of Hard Samples: A Data-Centric Approach to Synthetic Data for Medical Tasks (ICML 2025 · similarity 0.83)
- Tab-Shapley: Identifying Top-k Tabular Data Quality Insights (AAAI 2025 · similarity 0.83)
- Unified Insights: Harnessing Multi-modal Data for Phenotype Imputation via View Decoupling (NeurIPS 2024 · 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.