Researchers from the Vector Institute, GEMINI, Unity Health, and Diabetes Action Canada have published a model called CRISPNAM-FG that predicts the risk of diabetic foot complications in patients discharged from hospital while making each feature’s contribution to that prediction visible to clinicians.
The model’s full name is Competing Risks Survival Prediction using Neural Additive Models: Fine-Gray. According to the Vector Institute’s description of the work, it addresses a longstanding tension in clinical AI: recent deep learning survival models such as DeepHit and Neural Fine-Gray achieve strong predictive performance, but function as black boxes that clinicians cannot inspect or validate against domain knowledge.
CRISPNAM-FG uses a Neural Additive Model architecture in which each clinical variable is processed by its own small neural network, called a FeatureNet. Each FeatureNet produces a non-linear transformation of a single input feature, and those outputs are combined under the Fine-Gray framework, which models competing risks — in this case, diabetic foot complications versus in-hospital mortality from other causes. Unlike traditional survival approaches that remove patients from analysis once any competing event occurs, the Fine-Gray formulation continues tracking patients, according to the institute’s summary.
The model generates what the researchers call shape functions: curves showing how different values of a given variable — blood glucose, age, HbA1c, blood pressure — influence risk for each competing outcome. It also produces feature importance rankings identifying which factors most strongly drive each outcome.
The team validated the model on a cohort of 107,386 adult patients with diabetes discharged from 29 Ontario hospitals between April 2016 and March 2023. The data came from GEMINI, a hospital data platform. Foot complication was defined as a first subsequent hospitalization for foot ulcer, infection, gangrene, or amputation. The cohort had a median age of 72.0 years, and 46.2 percent were female, according to the source.
The institute’s description contrasts CRISPNAM-FG with post-hoc explanation methods such as LIME and SHAP, which the researchers argue “suffer from low fidelity” and “can be easily fooled.” Rather than applying an explanation layer after the fact, CRISPNAM-FG is designed so that interpretability is built into the architecture rather than added afterward.
The source does not include a direct comparison of CRISPNAM-FG’s predictive accuracy against DeepHit or Neural Fine-Gray on the Ontario cohort. The full paper is accessible via the Vector Institute page.
The collaboration between Vector, GEMINI, Unity Health, and Diabetes Action Canada reflects a model common in Canadian health AI research, where methodological expertise from the Vector network is paired with large-scale clinical datasets held by hospital networks. Unity Health Toronto operates St. Michael’s Hospital, among others, and has an established AI research program.