Remote assessment of Parkinson’s disease symptom severity based on group interaction feature assistance

Telemonitoring is an effective way to assess the severity of Parkinson's disease (PD). Due to heterogeneity and small sample sizes, the multi-task learning is applied to build the specific model for PD patients and prevent overfitting. However, the existing multi-task learning methods don't consider the nonlinear interaction between patients. Therefore, to improve the performance of the patient-specific prediction model, this paper proposes a group interaction feature assistance method (GIFA) for remote assessment of PD symptom severity. First, GIFA employs the nonlinear bidirectional long short-term memory network to explore the correlation among patient groups. Next, the incremental association Markov boundary is adopted to select causal features from group interaction features obtained by the bidirectional long short-term memory network to reduce negative transfer. Finally, the causal interaction features learned by the incremental association Markov boundary are input into the patient-specific prediction model to assist disease assessment, which is conducive to increasing the complementary information and improving the prediction performance. Experiment results on the public Parkinson's telemonitoring and mPower voice datasets show that GIFA model outperforms the cited state-of-the-art comparison methods for predicting Parkinson's disease symptom severity.