Research On The Evaluation Method Of Essential Tremor Based On Personalized Human Behavior Recognition

Essential tremor(ET)is a progressive neurological disorder,typically occur in the upper limbs,which makes patients' daily life extraordinarily difficult,the assessment of ET plays a crucial role in the effective health management of patients.ET rating scales are wildly used in clinical practice,but the diagnosis is less reliable due to clinicians' knowledge and experience and cannot reflect the tremor fluctuations among the patient's daily activities.How to comprehensively and accurately evaluate the true tremor level of patients is a challenge in current research.Human activity recognition(HAR)has become an important task in the field of ubiquitous perception,especially in medical care and health applications,the main purpose is to extract knowledge from motion data obtained by wearable sensor devices.Recently,researchers have made great progress in essential tremor evaluation by using human activity recognition technology,but there are still three problems that need to be solved.First,most methods are not considering the background activity when assessing patients' tremor which ignoring the inconsistency of tremor levels under different background activities;Secondly,other research works use motion data from healthy people to train common activity recognition system,did not take into account that the movement patterns between tremor patients and healthy people are different,makes the model performed poorly in patients' activity recognition;Finally,most tremor evaluation works based on wearable devices are shallow models,and the accuracy of evaluation needs to be improved.To solve the problem of evaluating the tremor level without considering background activities of patients,proposed a deep forest-based activity recognition model,dForest-HAR,to recognize 6 background activities,comprehensively evaluate the tremor levels of each background activity and accurately evaluate the real tremor level of patients.Aiming at the problem that the common activity recognition model can not adapt to the motion mode of tremor patients,a common model personalization algorithm is proposed,which can adjust the common model through a small number of specific user data to generate a personalized model that adapts to the user behavior model.To improving the accuracy of tremor evaluation,a dual-input and dual-output model,eval-net,integrated with CNN and LSTM structures was proposed.The twodimensional activity tensor and triaxial acceleration time series obtained by the sensor data fusion algorithm and time series of triaxial acceleration were input into CNN and LSTM module respectively.The depth model can automatically extract the high dimension characteristics to improve evaluation accuracy.Experiments on open dataset and tremor dataset show that the proposed dForest-HAR combined with the common model personalization algorithm achieves the highest accuracy and F1 score compared with the comparison algorithms.The evaluation result produced by the Eval-net are significantly correlated with the FTMTRS ratings of two neurologists.