Research map / Time series / Causal and generative time series
Temporal Island Sparse Autoencoders for Interpreting Clinical Time-Series Models
Muhan Yeo, Bong Gyun Kang, HyunGi Kim, Sungroh Yoon
Read the paperCheck your idea against 137,217 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.
Introduces the Temporal Island Sparse Autoencoder (TI-SAE), a backbone-specific sparse autoencoder for interpreting deep clinical time-series models trained on EHR data.
Problem
Deep learning models on electronic health records are accurate but opaque; post-hoc attribution gives per-cell importance scores lacking temporal coherence, cross-variable structure, and reusability across patients.
Approach
A backbone-specific sparse autoencoder with temporal island structure that learns reusable, temporally coherent clinical concept features from a model's internal representations, as an alternative to per-cell attribution scores.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| use TI-SAE soft selection instead of integrated gradients | frozen clinical backbones · MIMIC-III/IV, Discrete backbone | AOPC: .060 -> .086 | Improved |
| use TI-SAE on STraTS backbone | STraTS · MIMIC-III/IV | AOPC: .071 (TSR) vs .016 | Hurt |
| add sparse variable routing (ablation) | TI-SAE · MIMIC-III/IV | AOPC: removal costs 58-75% of AOPC | Improved |
| add temporal islands and channel organization (ablation) | TI-SAE · MIMIC-III/IV | AOPC | No clear effect |
| add temporal islands and channel organization (ablation) | TI-SAE · MIMIC-III/IV | channel coherence (%): 36.9 -> 91.9 | Improved |
| train dictionary before selector instead of applying selection to raw hidden states (ablation) | TI-SAE · MIMIC-III/IV | AOPC: .038 -> .086 (Discrete) | Improved |
| use multiple islands per concept (M=4) instead of a single island (M=1) (ablation) | TI-SAE · MIMIC-III/IV | faithfulness and plausibility | Improved |
| use learned split-normal window instead of symmetric Gaussian window (ablation) | TI-SAE · MIMIC-III/IV | faithfulness and plausibility | No clear effect |
| transfer MIMIC-III dictionary to MIMIC-IV without retraining | TI-SAE · MIMIC-IV (2014+) | AOPC: .149 -> .130 | No clear effect |
| remove dictionary labels from the verbalizer (ablation) | TI-SAE · verbalized explanations | LLM judge win rate | Hurt |
| use TI-SAE concept explanations instead of IG attribution | TI-SAE · verbalized explanations, clinicians | concept-level reasoning wins: 39-4 (7 ties) | Improved |
| use TI-SAE concept explanations instead of IG attribution | TI-SAE · verbalized explanations, clinicians | faithfulness wins: 4-4 (42 ties) | No clear effect |
Experimental setup
Baselines: post-hoc attribution methods
Builds on: sparse autoencoders
Closest papers by meaning
- Risk Horizons: Structured Hypothesis Spaces for Longitudinal Clinical Prediction (NeurIPS 2026 · similarity 0.85)
- Integrating Sequence and Image Modeling in Irregular Medical Time Series Through Self-Supervised Learning (AAAI 2025 · similarity 0.85)
- MOTOR: A Time-to-Event Foundation Model For Structured Medical Records (ICLR 2024 · similarity 0.83)
- LLM Interpretability with Identifiable Temporal-Instantaneous Representation (NeurIPS 2025 · similarity 0.83)
- Time-to-Event Pretraining for 3D Medical Imaging (ICLR 2025 · similarity 0.83)
- Autoregressive Sequence Modeling for 3D Medical Image Representation (AAAI 2025 · similarity 0.83)
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.
