Development, validation, and pilot implementation of a prognostic model to predict future foot complications among people with diabetes recently discharged from hospital.

OBJECTIVES

Many people with diabetes do not consistently receive guideline-recommended preventative foot care, increasing their risk for severe ulcers and amputations. Most people who undergo a lower-extremity amputation are hospitalized within the year prior. We leveraged hospitalization as a catalyst for targeted referral to community-based preventative foot care, using this clinical example to demonstrate prognostic model evaluation in multi-institutional hospital systems.

METHODS

We developed and validated prognostic models to predict the 1-year probability of a future foot complication (ulcer, infection, gangrene, or toe/foot/leg amputation) among people with diabetes discharged from a General Internal Medicine service. The data source included all General Internal Medicine discharges across 29 hospitals in Ontario, Canada (2016-2023). Predictors were prespecified based on literature and clinical judgment. We fit a Fine-Gray regression and a random survival forest for competing risks and evaluated performance using internal-external cross-validation. Model evaluations assessed both statistical measures (discrimination, calibration) and clinical utility (net benefit) across held-out hospitals. The better-performing model was then piloted at one participating hospital (Jan-Mar, 2025) to prioritize highest-risk patients for referral to a community-based preventative foot care clinic after discharge home. Reporting adheres to Transparent Reporting of a Multivariable Prediction Model for Individual Prognosis or Diagnosis (TRIPOD)-Cluster standards.

RESULTS

Among 107,830 index hospitalizations, 2.7% (N = 2912) developed a foot complication, and 12.1% (N = 13,001) died within 1 year of discharge. Both models demonstrated good discrimination in held-out validation hospitals (pooled area under the receiver operating characteristic curve [AUROC]: 0.71, 95% confidence interval [CI]: 0.70-0.73; 0.72, 95% CI: 0.70-0.73). However, discrimination and calibration varied substantially across hospitals. The Fine-Gray model produced better calibration, including among high-risk patients and key clinical subgroups. Using the Fine-Gray model to make prospective predictions, 160 consecutive eligible patients had a 1-year median outcome risk of 1.1% (25th-75th, 0.7%-1.9%; range: 0.2%-14.0%). Among the 107 patients who could be reached by phone, 44 (41%) accepted referral to the community-based preventative foot care clinic.

CONCLUSION

Through this example we illustrate how a learning health system can leverage hospital encounters for community-based preventative care and demonstrate modern methods for prognostic model evaluation that reveal substantial variation in performance across hospitals.