Predicting diabetic foot amputation using machine learning and statistical models: a prospective cohort study

Abstract

Objective: To compare the predictive performance of an artificial neural network model with statistical logistic regression model for predicting the risk of diabetic foot amputation.

Method: The prospective cohort study was conducted at the Dow University Hospital, Karachi, between October 2024 and June 2025 and adult comprised patients with diabetic foot ulcers. The patients were monitored for the occurrence of both minor and major diabetic foot amputations. Demographic, clinical and laboratory data was collected at baseline. Artificial neural network and logistic regression models were created to evaluate the likelihood of diabetic foot amputation. Data were analysed using R version 4.5.0, and model performance was evaluated using appropriate performance measures.

Results: Of the 496 patients, 93(18.75%) underwent diabetic foot amputation; 63(67.74%) males and 30(32.26%) females with mean age 67.36+0.36 years. The remaining 403(81.25%) patients did not undergo diabetic foot amputation; 243(60.30%) males and 160(39.70%) females with mean age 51.29+2.19 years. Among those who underwent amputation, 82(88.17%) had minor amputations and 11(11.83%) had major amputations. Amputation was more common in male patients 63(67.74%) than in female patients 30(32.26%). The artificial neural network ANN model outperformed the logistic regression model with an area under the receiver operating characteristic curve 0.862, accuracy 86.0%, and 95% confidence interval 0.791-0.892 (p<0.001). The corresponding values for the logistic regression model were 0.726, 71.2% and 0.651-0.756 (p=0.042).

Conclusion: Based on empirical findings, the artificial neural network ANN model was found to be a useful tool for predicting the risk of diabetic foot amputation, outperforming the classic logistic regression model.

Key Words: Diabetic foot, Amputation, Machine learning, Prognosis, Logistic models.