---
title: "Temporal Island Sparse Autoencoders for Interpreting Clinical Time-Series Models"
source: "arXiv AI/CL/LG（论文）"
source_tier: "t2"
url: "https://arxiv.org/abs/2611.03420v1"
permalink: "https://feed.astralor.com/items/itm_6b51a6fd734b180d15391883"
published: "2026-11-02T16:43:03.000Z"
score: 57
channel: "ai"
tags: ["Temporal Island Sparse","Autoencoders","Interpreting Clinical Time-Series","Models Deep","Post-hoc","Deep"]
---

# Temporal Island Sparse Autoencoders for Interpreting Clinical Time-Series Models

## At a glance

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 im…

## Key points

- 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.

## Full text

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.

---

Verify against the original before citing; curated by Astral Feed: https://feed.astralor.com/items/itm_6b51a6fd734b180d15391883