Using Deep Learning to Detect Motor Impairment in Early Parkinson's Disease from Touchscreen Typing

Assessing older adults' motor control ability is crucial for early diagnosis of Parkinson's Disease (PD). In this paper, we investigate how to use deep learning to detect motor impairment of older adults via analyzing touchscreen typing data, which could result in the early diagnosis of PD. Our investigation shows that deep learning is promising in analyzing touchscreen typing data. Among the four deep learning models (LSTM, LSTM-CNN, CNN-LSTM, and 1D CNN), LSTM-CNN yields the best performance. On a 102-subject dataset [53], LSTM-CNN achieved an AUC of 0.95 and an F1-score of 0.90 in leave-one-out PD classification, improving the performance of previously used SVM method (AUC = 0.88, F1-score = 0.73) [53]. LSTM-CNN also performed well on an in-the-clinic typing dataset [27] (AUC = 0.86, F1-score = 0.86), and significantly improved the F1-score of the previously proposed 1D CNN method (AUC = 0.89, F1 = 0.80). The promising performance of the LSTM-CNN model can also be generalized to other touchscreen interactions including flick, drag, handwriting, and pinch. It achieved better performance than the previous SVM method [53]. Our research showed that deep learning is effective in detecting early PD's motor symptoms via analyzing smartphone interaction data, and the proposed LSTM-CNN model is a promising neural network structure for performing such analysis. Overall, our research advances the understanding of how to assess the motor control ability of older adults via smartphone interactions.