Intelligent Evaluation Of Finger Flexibility In Patients With PD Based On Multi-source Information

Artificial intelligence(AI)technology has become widely used in the medical field to assist in the intelligent diagnosis and treatment of diseases,including Parkinson’s disease.Both in the diagnosis stage and treatment stage,the disease requires significant effort from clinicians,patients,and their families.Machine learning and deep learning algorithms can predict the state of patients with Parkinson’s disease,helping clinicians save time and provide valuable references.Such AI models also have practical significance for enabling patients to monitor themselves at home.This paper collected 74 sample data from the Neurology Department of the First Hospital of Shanxi Medical University,and used computer vision and multi-source data fusion methods to study the hand flexibility testing of Parkinson’s patients.The main work is as follows:1.The paper collected RGB video data,and cropped video containing only hands.It firstly trained a autoencoder to extract prior knowledge features of video streams and verified its effectiveness.Next,it used data augmentation methods to expand the sample size and constructed a LSTM model to study the effect of different sample lengths on score prediction.The results showed that samples with a length of 4 seconds had the best performance,achieving an accuracy rate of 80.95% and a recall rate of 84.16%.2.The paper also used the Media Pipe Hands algorithm to extract the skeleton outline of the hand from the clipped hand motion video,and obtained the three-dimensional coordinates of the key points of the finger.Then it extracted detail features from the trajectory,including motion features based on the scoring rules in the MDS-UPDRS scale and time series features,and used feature filtering and feature selection to reduce dimensions.Finally,it used a random forest model for prediction,the accuracy rate of the two classification model can reach 89.29%,and the recall rate of the abnormal class can reach 96.25%,which is better than the multi classification model.3.This paper integrated muti-source data and constructed a hand flexibility score prediction model that fuses video modal features and sensor data features.According to these results,multi-source data model had better performance compared with single-modal models,and early fusion helps a lot than late fusion.The random forest model based on skeleton features and sensor data features had the best performance,and its accuracy,recall,precision and f1-score all reaching over 90%.As the data in this article was collected in an open experimental environment,which is closer to real life,the results had certain practical significance.