Machine learning approach to diabetic foot risk classification with biomechanics data

Abstract

Diabetic foot wound onset, recurrence and healing are globally considered a challenge [1]. In particular, the prevention topic is pursued through several recommendations, however, it is desirable to classify the risk of ulceration before their occurrence, to prevent severe ulcers and subsequent amputations [1]. With regard to the biomechanical properties of the diabetic foot, diabetic patients showed higher load on the push-off associated with lower dorsi-plantar flexion of the ankle during stance and the push-off phase [2]. The aim of this study was to automatically define a risk index of the diabetic foot ulceration by applying machine learning (ML) algorithms to a dataset of standard gait analysis, musculoskeletal modelling and finite element modeling data and of routine diabetic neuropathy screening scores [4]. The study included retrospectively the data of 80 subjects (26.3% healthy subjects, 37.5% diabetics without neuropathy and 36.3% diabetics with neuropathy). Diabetic patients were clinically assessed according to the ADA guidelines [3] as reported in [4]. Kinematic, kinetic and electromyographic (EMG) data were acquired through a modified version of the IORgait protocol [3,5] while subjects were walking several gait cycles at self selected speed. Muscle forces were estimated with the Opensim gait 2392 model and internal stresses were determined using a foot finite element model [6]. Joint angles and moments, ground reaction forces, plantar pressures, EMG envelopes and timing of activation, muscular forces and internal stresses were considered for the ML algorithm. For what regards the clinical features, presence or absence of peripheral and autonomic neuropathy, retinopathy, microalbuminuria, vasculopathy, knee-hallux-toes deformities, plantar callosity were selected for the ML algorithm. MongoDB was chosen as database program and data extractions were performed through the library Pandas. Regarding the ML algorithm, the values of different gaits were merged using an arithmetic mean, in order to obtain a narrower dataset and more stable values of the features. Due to the size of the dataset, ‘simple’ and linear models were adopted, and, among these, Logistic Regression, Linear Model and Perceptron were chosen. For completeness, some more complex classifiers like Kernel Support Vector Machine, Random Forest, Gradient Boosting Tree and eXtreme Gradient Boosting Tree were also adopted but showed lower performances. Logistic Regression with the adoption of 4 features achieved the best result: 83% of precision and 83% of accuracy (Fig. 1). Once the model was developed, a React Native mobile application was implemented with a relative Application Protocol Interface (API) to produce an environment to host and exploit the model. Diabetic foot risk classification according to ADA [3] was used as comparison for the results of the ML approach. The performance of the model is not considered to be optimal, but the results are encouraging from a diabetic foot risk point of view. Future development includes the expansion of databases and the exploration of other ML algorithms.