Interpretable Deep Survival Model Predicts Diabetic Foot Complications in Ontario

Global: Interpretable Deep Survival Model for Diabetic Foot Complications

Researchers led by Dhanesh Ramachandram and Anne Loefler have introduced a new deep learning survival model, named CRISPNAM‑FG, designed to predict post‑discharge foot complications among diabetic patients treated at 29 Ontario hospitals between 2016 and 2023. The model, described in an arXiv preprint submitted on 16 Nov 2025, aims to combine high predictive accuracy with intrinsic interpretability for competing‑risk analysis.

Model Architecture and Methodology

CRISPNAM‑FG builds on Neural Additive Models (NAMs) by assigning separate projection vectors to each competing risk, enabling the direct estimation of the Cumulative Incidence Function through the Fine‑Gray formulation. This structure allows the model to generate transparent shape functions and feature‑importance visualizations for each risk factor.

Validation on Benchmark Datasets

The authors report that the model was evaluated on several established benchmark survival datasets, where it achieved performance comparable to state‑of‑the‑art deep survival models while retaining interpretability.

Application to Diabetic Foot Complications

Applying CRISPNAM‑FG to the Ontario cohort, the study forecasts the likelihood of foot‑related adverse events after hospital discharge. The analysis leverages demographic, clinical, and treatment variables collected across the province’s health system.

Implications for Clinical Practice

According to the authors, the model’s transparent predictions could facilitate clinician trust and support decision‑making in high‑risk diabetic care, addressing concerns about the “black‑box” nature of many AI‑driven prognostic tools.

Future Directions

The paper suggests further work to integrate the model into electronic health‑record workflows and to assess its performance prospectively in real‑time clinical settings.

This report is based on information from arXiv, licensed under Academic Preprint / Open Access. Based on the abstract of the research paper. Full text available via ArXiv.