Explainable machine learning approach on biomechanical features to identify weakness in a population-based setting on aging

Abstract

Weakness, as measured by maximal Hand Grip Strength (HGS), represents one of the five criteria used in Fried's definition of frailty [1] and is associated with a wide range of health conditions, which makes it challenging to delineate what body system processes are responsible for weakness. [2]. Still, poor studies have investigated the associations between HGS and dynamic functional assessments [3]. To identify a pattern of functional characteristics, extracted from the 5-repetitions-sit-to-stand (5STS) test biomechanical signals best predict weakness. An Explanation approach to a Machine Learning model was also used. In a subcohort of the longitudinal study of aging [4], 86 subjects over 65 performed the 5STS test [5,6]. They were equipped with an IMU on the L5 vertebra and four sEMG probes (BTS Bioengineering) on the Gastrocnemius Medialis and Tibialis Anterior both side muscles. Several kinematic and muscular features were extracted from the cycle, standing and sitting phases. A handgrip dynamometer was used to measure HGS. Men and women who were considered weak had HGS<26 kg and <16 kg, respectively. Socio-demographic information (age, sex and BMI) was also included. The final dataset consisted of 119 features for all subjects. We first performed the undersampling of the majority class (without weakness); then the dataset was divided into 70% training and 30% testing and normalised using the z-score method. Because of the curse of dimensionality, a pipeline for feature selection and hyperparameter tuning, using the GridSearchCV method, was defined to obtain the best Kernel-SVM model. The best model was chosen according to the accuracy score. To evaluate our model accuracy, precision and recall were calculated. All the analyses were performed using the Scikit-Learn library [7] with Python 3.6. To explain our model Python's SHAP library was used [8]. From the hyperparameter tuning, we obtained six features: hip power (Whip), power along the vertical axis (Wvert), and cycle jerk along the vertical axis and its coefficient of variation, age, and sex. Fig. 1 shows the boxplots for the biomechanical selected variables.The model showed 90.0% and 85.7% accuracy on the training and testing sets, respectively. The precision of 100%, recall of 71%, and f1-score 83%, while the precision of 78%, recall of 100%, and f1-score of 88% was obtained on the class without weakness and its counterpart, respectively.The explainability analysis showed that age, Wvert and Whip were the three most important variables in predicting weakness in absolute terms. Sex resulted being the least important variable. Picture 1 - "Boxplot of the biomechanical selected features according to the weakness condition"Download : Download high-res image (71KB)Download : Download full-size image Measures of HGS are associated with deficits in several physical functions. In a population-based setting, we identified biomechanical features from 5STS related to stability that could help predict weakness in a free-living environment.