How AI Turns Messy EHR Into Clear Survival Predictions

Sep 8, 2025

26:03

forum

Ask episode

view_agenda

Chapters

auto_awesome

Transcript

info_circle

Episode notes

Can AI forecast ICU risk from the first 36 hours of EHR data?

University of Washington researcher Sihan explains TrajSurv, a survival-prediction model that converts noisy, irregular ICU time series into interpretable latent trajectories using Neural Controlled Differential Equations (NCDEs) and time-aware contrastive learning aligned to SOFA. We cover how trajectories outperform snapshots, handle missingness without heavy imputation, and remain clinically legible via vector-field feature importance and trajectory clustering.

Validated on MIMIC-III and eICU with reported C-index ≈0.80 and cross-cohort ≈0.76, TrajSurv points to safer escalation, de-escalation, and bed allocation in the ICU.

In this episode: survival prediction basics; limits of Cox/RSF vs deep time-series models; NCDE explained in plain language; first-36h feature set (53 labs/vitals/demographics); metrics (C-index, Brier, dynamic AUC); interpretable clustering linked to outcomes; and what’s next—adding interventions for counterfactual simulation and extending to oncology.

Link to the paper: https://arxiv.org/abs/2508.00657

Timestamps

00:00 Why trajectories beat snapshots in EHR

01:00 Guest intro: Sihan, UW Biomedical Informatics

01:40 Survival prediction 101 and clinical use

03:40 From Cox/RSF to deep learning on time-varying data

05:03 What is TrajSurv (pronounced “traj-surf”)?

06:16 NCDE explained with the “ship + weather” analogy

08:14 Handling irregular sampling and missing data

09:14 Time-aware contrastive learning aligned to SOFA

10:47 Datasets: MIMIC-III and eICU; first 36h features (labs, vitals, demo)

12:40 Results: C-index ≈0.80; cross-cohort ≈0.76; interpretability