Establishment and Reliability Evaluation of Prognostic Models in Diabetic Foot

To study the risk factors affecting amputation and survival in patients with diabetic foot (DF) and to construct a predictive model using the machine learning technique for DF foot amputation and survival and evaluate its effectiveness.

A total of 200 patients with DF hospitalized in the First Affiliated Hospital of Shantou University Medical College in China were selected via cluster analysis screening, Kaplan-Meier survival calculation, amputation rate and Cox proportional hazards model investigation of risk factors associated with amputation and death. In addition, we constructed various models, including Cox proportional hazards regression analysis, the deep learning method convolution neural network (CNN) model, backpropagation (BP) neural network model, and backpropagation neural network prediction model after optimizing the genetic algorithm. The accuracy of the 4 prediction models for survival and amputation was assessed, and we evaluated the reliability of these computational models based on the size of the area under the ROC curve (AUC), sensitivity and specificity.

We found that the 1-year survival rate in patients with DF was 88.5%, and the 1-year amputation rate was 12.5%. Wagner's Classification of Diabetic Foot Ulcers grade, ankle-brachial index (ABI), low-density lipoprotein (LDL), and percutaneous oxygen partial pressure (TcPO2) were independent risk factors for amputation in patients with DF, while cerebrovascular disease, Sudoscan sweat gland function score, glycated hemoglobin (HbA1c) and peripheral artery disease (PAD) were independent risk factors for death in patients with DF. In addition, our results showed that in the case of amputation, the COX regression predictive model revealed an AUC of 0.788, sensitivity of 74.1% and specificity of 83.6%. The BP neural network predictive model identified an AUC of 0.874, sensitivity of 87.0% and specificity of 87.7%. An AUC of 0.909, sensitivity of 90.7% and specificity of 91.1% were found after optimizing the BP neural network prediction model via genetic algorithm. In the deep learning CNN model, the AUC, sensitivity and specificity were 0.939, 92.6%, and 95.2%, respectively. In the analysis of risk factors for death, the COX regression predictive model identified the AUC, sensitivity and specificity as 0.800, 74.1% and 85.9%, respectively. The BP neural network predictive model revealed an AUC, sensitivity and specificity of 0.937, 93.1% and 94.4%, respectively. Genetic algorithm-based optimization of the BP neural network predictive model identified an AUC, sensitivity and specificity of 0.932, 91.4% and 95.1%, respectively. The deep learning CNN model found the AUC, sensitivity and specificity to be 0.861, 82.8% and 89.4%, respectively.

To identify risk factors for death, the BP neural network predictive model and genetic algorithm-based optimizing BP neural network predictive model have higher sensitivity and specificity than the deep learning method CNN predictive model and COX regression analysis.