Identification and Interpretation of Gait Analysis Features and Foot Condition by Explainable AI

Abstract Background Clinical gait analysis is a crucial step for identifying foot disorders and surgery planning. However, a large amount of gait data makes this assessment difficult and time-consuming. There are separate efforts to reduce its complexity by manually or automatically finding features (e.g. minimum of a joint angle in a specific axis), identifying the foot condition by Machine Learning (ML), and interpreting the outcome by explainable artificial intelligence (xAI). Methods In this article, we explore the potential of state-of-the-art ML algorithms to automate all these steps for a set of 6 foot conditions. New features are created manually and then recursive feature elimination is employed based on Support Vector Machines (SVM) and Random Forest (RF) to eliminate the features with low variance. SVM, RF, K-nearest Neighbor (KNN), Logistic Regression (LREGR), and Majority Voting (MV) algorithms are compared for classification and Local Interpretable Model-agnostic Explanation (LIME) is used for the interpretation of the outcome of the ML models. 40 features are eliminated and 334 features are given to the classifier models as inputs. Results The foot conditions are classified with a maximum average accuracy of 0.86 by KNN and MV, maximum average recall of 0.97 by KNN, and max average F1 score of 0.86 by KNN and MV. Conclusions High success scores indicate that the relation between the selected features and foot conditions should be strong and meaningful, potentially indicating clinical relevance. All models are interpreted for each foot condition for random 20 patients and the most contributing features are graphically demonstrated. The proposed ML pipeline can be easily extended for other foot conditions and retrained as new data arrives. It can help experts and physicians in the identification of foot conditions and the planning of potential surgeries.