Discriminating Between Marijuana and Alcohol Gait Impairments Using Tile CNN With TICA Pooling

Goal: To investigate whether machine learning analyses of smartphone sensor data can discriminate whether a subject consumed alcohol or marijuana from their gait. Methods: Using first-of-a-kind impaired gait datasets, we propose MariaGait, a novel deep learning approach to distinguish between marijuana and alcohol impairment. Subjects' time-series smartphone accelerometer and gyroscope sensor gait data are first encoded into Gramian Angular Field (GAF) images that are then classified using a tiled Convolutional Neural Network (CNN) with TICA pooling. To mitigate the insufficiency of positively labeled alcohol and marijuana instances, the tiled CNN was pre-trained on sober gait samples that were more abundant. Results: MariaGait achieved an accuracy of 94.61%, F1 score of 88.61%, and 94.33% ROC AUC score in classifying whether the subject consumed alcohol or marijuana, outperforming baseline models including Multi-Layer Perceptron (MLP), Long Short Term Memory (LSTM), Multi-head CNN and Multi-head LSTM, Random Forest and Support Vector Machines (SVM)). Conclusions: Our results demonstrate that MariaGait could be a practical, non-invasive approach to determine which substance a subject is impaired by from their gait.