A Machine Learning Framework for Activity Recognition with Robotic Ankle Foot Orthoses

Novel auto-adjusting ankle foot orthoses (AFOs) have been developed for individuals with cerebral palsy, stroke, or traumatic brain injury. These devices remedy the issue of poor adaptability in fixed stiffness AFOs. However, new controllers are needed to make robotic AFOs adaptable to different walking terrain and tasks. We embedded two machine learning models, Long-Short Term Memory (LSTM) and Random Forest (RF), in a robotic, adjustable AFO for stair ambulation classification. Pilot testing determined the minimum training data size to achieve F1 scores of 0.99 was 60 steps per class. Validation was conducted in six individuals with cerebral palsy (GMFCS I-II, Age 13-38) during training and testing periods that occurred outside standard lab environments. On average, the personalized RF models had 9% better accuracy and 34x faster speed than the personalized LSTM models. Personalized models yield 6-11% higher F1 scores than leave-one-out generalized models. These results demonstrate the feasibility for machine learning algorithms to be used in real-world activity recognition for an auto-adjusting AFO controller.