Classification of mild and severe adolescent idiopathic scoliosis (AIS) from healthy subjects via a supervised learning model based on electromyogram and ground reaction force data during gait

At early stages, adolescent idiopathic scoliosis (AIS) is quite hard to be distinguished from healthy (HC) subjects by the naked eye. AIS demands multiple corrective surgeries when detected later, thereby causing significant physical and psychological trauma as no mathematical models exist for the classification of mild AIS (MS) (20° < Cobb’s angle < 40°) from HC, we propose a k-nearest neighbour (kNN) method based model. In this work, we collected both the EMG and GRF data from nine severe AIS (SS), three MS and four female HC during gait. Delayed muscle activation in Erector spinatus Iliocostalis, Gluteus Medius and Gastrocnemius lateralis was observed in SS compared to HC. However, no such distinction was noticed between MS and HC motivating for a mathematical model. Eighteen time-domain and nine frequency-domain features were computed from the EMG data of 14 lower extremity muscles, while five time-domain features were calculated from GRF data during gait. Out of all the features computed for each subject, the principal component analysis (PCA) yielded 15 principal components that coupled both time and frequency domains (TFD). Further, the kNN model classified SS, MS and HC from each other by these 15 TFD features. The model was trained and validated using 32 and 21 EMG and GRF data datasets during gait, respectively. The classification and validation accuracy of 90.6% and 85.7% were obtained among SS, MS and HC. The proposed model is capable of early detection of AIS and can be used by medical professionals to plan treatments and corrective measures.