Logo
ScienceStack
Table of Contents
Fetching ...
Sign In

Interpretable Fine-Gray Deep Survival Model for Competing Risks: Predicting Post-Discharge Foot Complications for Diabetic Patients in Ontario

Dhanesh Ramachandram, Anne Loefler, Surain Roberts, Amol Verma, Maia Norman, Fahad Razak, Conrad Pow, Charles de Mestral

TL;DR

CRISPNAM-FG addresses interpretability in competing risks survival analysis by embedding Neural Additive Models within the Fine-Gray CIF framework to produce additive, risk-specific hazard contributions. The model learns per-feature, per-risk projections with unit-norm constraints, enabling direct comparison across risks while providing shape-function visuals for interpretation. Across benchmark datasets and the GEMINI diabetes cohort in Ontario, CRISPNAM-FG demonstrates strong discrimination (TD-AUC/TD-CI) and transparent explanations, though calibration can lag behind some baselines. This work advances safe, auditable survival prediction in healthcare by delivering interpretable CIF estimates alongside competitive predictive performance, with potential for clinical deployment and further methodological refinements.

Abstract

Model interpretability is crucial for establishing AI safety and clinician trust in medical applications for example, in survival modelling with competing risks. Recent deep learning models have attained very good predictive performance but their limited transparency, being black-box models, hinders their integration into clinical practice. To address this gap, we propose an intrinsically interpretable survival model called CRISPNAM-FG. Leveraging the structure of Neural Additive Models (NAMs) with separate projection vectors for each risk, our approach predicts the Cumulative Incidence Function using the Fine-Gray formulation, achieving high predictive power with intrinsically transparent and auditable predictions. We validated the model on several benchmark datasets and applied our model to predict future foot complications in diabetic patients across 29 Ontario hospitals (2016-2023). Our method achieves competitive performance compared to other deep survival models while providing transparency through shape functions and feature importance plots.

Interpretable Fine-Gray Deep Survival Model for Competing Risks: Predicting Post-Discharge Foot Complications for Diabetic Patients in Ontario

TL;DR

CRISPNAM-FG addresses interpretability in competing risks survival analysis by embedding Neural Additive Models within the Fine-Gray CIF framework to produce additive, risk-specific hazard contributions. The model learns per-feature, per-risk projections with unit-norm constraints, enabling direct comparison across risks while providing shape-function visuals for interpretation. Across benchmark datasets and the GEMINI diabetes cohort in Ontario, CRISPNAM-FG demonstrates strong discrimination (TD-AUC/TD-CI) and transparent explanations, though calibration can lag behind some baselines. This work advances safe, auditable survival prediction in healthcare by delivering interpretable CIF estimates alongside competitive predictive performance, with potential for clinical deployment and further methodological refinements.

Abstract

Model interpretability is crucial for establishing AI safety and clinician trust in medical applications for example, in survival modelling with competing risks. Recent deep learning models have attained very good predictive performance but their limited transparency, being black-box models, hinders their integration into clinical practice. To address this gap, we propose an intrinsically interpretable survival model called CRISPNAM-FG. Leveraging the structure of Neural Additive Models (NAMs) with separate projection vectors for each risk, our approach predicts the Cumulative Incidence Function using the Fine-Gray formulation, achieving high predictive power with intrinsically transparent and auditable predictions. We validated the model on several benchmark datasets and applied our model to predict future foot complications in diabetic patients across 29 Ontario hospitals (2016-2023). Our method achieves competitive performance compared to other deep survival models while providing transparency through shape functions and feature importance plots.

Paper Structure

This paper contains 26 sections, 18 equations, 4 figures, 5 tables.

Table of Contents

  1. Introduction
  2. Related Work
  3. Methodology
  4. Competing Risks in Survival Analysis
  5. The Cause-Specific Hazard Method
  6. Fine-Gray Subdistribution Method
  7. Model Architecture
  8. Neural Additive Model (FeatureNet)
  9. Risk-Specific Projections
  10. Additive Risk Aggregation
  11. Fine-Gray Subdistribution Hazards
  12. Fine-Gray Partial Likelihood Loss
  13. Cumulative Incidence Function Estimation
  14. Interpretability Mechanisms
  15. Experiments
  16. ...and 11 more sections

Figures (4)

  • Figure 1: CRISPNAM-FG Architecture Diagram
  • Figure 2: Cohort Generation Flow
  • Figure 3: Shape functions generated by CRISPNAM-FG for the GEMINI Foot Complication dataset, averaged across folds. Ten features per risk are shown, randomly selected from the 20 most important.
  • Figure 4: Feature Importances computed by CRISPNAM-FG for Competing Risks for the GEMINI Dataset