Introduction:
Diabetic foot patients showed poor prognosis in the overall 5-year survival rate. The survival rate was even worse after major limb amputation. Several risk factors have been identified such as age, BMI ratio, renal function, peripheral artery diseases, nutrition, and so on. Blood sugar and underlying comorbidity control, revascularization for occlusive vessels, and wound debridement were the main management strategies for these patients. On the other hand, previous research showed that early amputation may provide a better outcome than late amputation. In our cohort, we found some patients suffered from major limb amputation despite receiving multiple times of debridement. In order to avoid unnecessary debridement, we have designed a diagnostic tool with machine learning that could predict the amputation risk and prognosis of diabetic foot patients.

Materials and Methods:
All the information and records of diabetes foot patients, including age, gender, BMI ratio, initial wound condition, diabetes treatment, diabetes mellitus-related complications, outcome, discharge condition, and laboratory and examination data from January 2018 to December 2020, have been collected retrospectively. During the machine learning model building, data had been preprocessed by variable calculation, data balancing, and normalization. SMOTE technique had been applied for data balancing and Z score standardization had been used for data normalization. 70% of data had been used for data training and 30% of data was entered for validation. Model performance evaluation was performed by accuracy value, ROC curve, Precision-Recall curve, and Cohen’s Kappa coefficient. Amputation risk was set as the major outcome, and each risk factor's influence was measured as the secondary outcome.

Result :
From January 2018 to December 2020, information and records of 1241 diabetes foot patients were enrolled for data processing. Four algorithms have proceeded and Random Forest showed an accuracy of 97.931% which was the best prediction accuracy in all algorithms. It also presented good accuracy and prediction ability in the model evaluation by the ROC curve and Precision-Recall curve (Precision: 0.986, Recall: 0.972, F1 score: 0.979). Further analysis by SHAP multi-class summary plot showed that C-reactive protein value, diabetes mellitus-related neuropathy (ANS), white blood cell count, blood creatinine value, and patient under intermittent hemodialysis were the most influential factors in the prediction of major limb amputation.

Conclusion :
We have provided an applicable and precise model for predicting major limb amputation in diabetes mellitus foot patients which could prevent those patients from receiving non-beneficial debridement surgeries so that reduced the operative risks and hospital stay and also improved the prognosis at the same time. Moreover, initial infection status and renal function revealed a significant influence on the risk of major limb amputation in those diabetes mellitus foot patients.

Download