Early Prediction Of Freezing Of Gait In Parkinson’s Disease Based On Deep Learning

Freezing of gait(FOG)is one of the late-stage gait disorders of Parkinson’s disease,which often leads to falls,injuries,and even death.Therefore,the prediction and warning of FOG are of great clinical significance for preventing patient falls.Although previous studies have shown that predicting FOG using wearable sensors is feasible,there are still some shortcomings: most deep learning methods only detect FOG rather than predict it;it is unable to accurately predict all FOG,resulting in some FOG occurring without warning;the warning time is too short compared to the occurrence time of FOG,which leads to patients falling before they have enough reaction time.To solve these problems,this thesis proposes a new deep learning model that can predict and provide warnings for FOG at least 2 seconds in advance,giving patients enough reaction time.The main work is as follows:(1)This thesis uses three modal signals: electroencephalography(EEG),electromyography(EMG),and acceleration/gyroscope(ACC/GYRO)to extract physiological signal features during normal walking and the transition period 2 seconds before FOG occurs.Statistical analysis showed that the majority of features of the three signals exhibited significant differences between normal walking and the transition period(p-value<0.05).Among them,the difference of ACC/GYRO signals between normal walking and the transition period was the most significant,therefore ACC/GYRO signal was selected for subsequent prediction work.(2)This thesis proposes an end-to-end deep learning model consisting of two structural parts.One part is a multi-scale convolution structure with one-dimensional and two-dimensional convolutional neural networks that capture the spatial relationship of sensor signals collected from various parts of the body.The other part is a bi-directional long short-term memory network model based on a Twin-tower Sequence Structure,consisting of two branches with the same architecture but without shared parameters,to extract temporal information of sensor signals.The model is used for FOG prediction with an accuracy of 85.2%,an F1 score of 85.5%,and an AUC(Area under Curve)of85.8%.(3)This thesis proposed a Knowledge Distillation(KD)framework.The teacher network model(TNM)uses all sensors,while the student network model(SNM)uses fewer sensors.SNM uses KD technology to learn valuable privileged information from TNM,reducing the number of sensors while improving the accuracy of FOG prediction.The results showed that the AUC of SNM with KD was 80.3%,while without KD it was75.2%.This thesis demonstrates the feasibility of using deep learning model for predicting FOG and improving prediction accuracy through KD on the basis of using the minimum number of sensors,thus indicating its potential in rehabilitation engineering.The study shows that there is a transition phase a few seconds before FOG occurs,during which there are differences in physiological signals compared to normal walking.FOG can cause changes in multiple physiological signals,with changes in kinematic signals being more pronounced than those in EEG and EMG,making kinematic signals more suitable for FOG research.Predicting brain dysfunction in advance using current body behavior information can help understand the information exchange between the brain and body.