Interpretable clinical concepts emerge from patient time series models
Temporal Island Sparse Autoencoders for Interpreting Clinical Time-Series Models
Machine Learning
Summary
Deep learning can predict medical outcomes from patient records but usually doesn't explain what medical ideas the model uses internally. The authors created a new method called Temporal Island Sparse Autoencoder (TI-SAE) that finds clear, meaningful clinical concepts shared across patients based on when and which health measurements matter. This approach makes it easier to understand the model’s decisions over time rather than just giving quick summaries of importance. They tested it on real hospital data and found it identifies structured patterns that reflect clinical thinking while still matching strong prediction accuracy.
What this means in practice
- •For hospital data teams: Improve clinical model explanations by extracting consistent, interpretable patient-level concepts across time from hospital records.
- •For healthcare software developers: Build tools that visualize meaningful clinical patterns driving predictions to support clinicians in understanding AI decision processes.$Commercial implications: Enables development of user-friendly interpretability modules in AI health products, improving trust and adoption by clinicians.
Authors
Muhan Yeo, Bong Gyun Kang, HyunGi Kim, Sungroh Yoon
Abstract
Deep learning models for electronic health records achieve strong predictive performance but provide little insight into what clinical concepts they represent internally. Post-hoc attribution methods produce per-cell importance scores that lack temporal coherence, cross-variable structure, and reusability across patients. We introduce the Temporal Island Sparse Autoencoder (TI-SAE), which learns a backbone-specific, patient-shared clinical concept dictionary from a frozen predictive model's representations. Each dictionary entry (latent) is defined by learned temporal Gaussian islands (specifying when the concept is active) and a sparse variable-composition vector (specifying which variables are involved). A decoupled selector then identifies which dictionary entries drive each patient's prediction, separating representation learning from attribution. We evaluate TI-SAE across five backbone architectures spanning different representation strategies on MIMIC-III and MIMIC-IV (three prediction tasks each). TI-SAE produces structured clinical concepts that go beyond transient attribution heatmaps while achieving strong performance on standard faithfulness benchmarks.