Study On Construction Of An Assessment Model For Adolescents’ Upper Cross Syndrome Based On Human Key Point Recognition Technology

Aim: To achieve an accurate assessment of the prevalence of Upper Crossing Syndrome in adolescents,this study identifies Upper Crossing Syndrome and uses Open Pose as a fully automated identification method to make a preliminary diagnosis of the student’s presence or absence of Upper Crossing Syndrome.Methods: In this study,students from three schools in Changchun are recruited as subjects and tested by measurement method for the presence of Upper Cross Syndrome.Differential analysis of variables is performed by using an independent sample t-test.This study uses Open Pose to detect key points on the human skeleton,extract the features needed to train an assessment model for Upper Crossing Syndrome,and perform feature selection based on a genetic algorithm.Using the confusion matrix to compare three different machine learning models(Extra Tree Classifier,Gradient Boosting Classifier,XGBoost Classifier)trained with the input of selected features and corresponding labels with three deep learning models(Res Net50,Alex Net,VGG-19)trained with the direct input of raw image data and corresponding labels to identify optimal model.Finally,the author conducts ablation experiments on selected features to verify their validity and to compare the effects of different feature combinations on the model.Results: 62%(1024)of students are diagnosed with Upper Crossing Syndrome;52.9%(542)of males and 47.1%(482)of females;38%(627)of students are diagnosed without Upper Crossing Syndrome,52.8%(331)of males and 47.2%(296)of females,with no significant difference between genders.The mean age,weight,and BMI of students with and without Upper Crossing Syndrome are not significantly different whereas height is significantly different(p<0.01);the mean angle1,angle2,angle3,and angle4 of students with Upper Crossing Syndrome are significantly lower than those of students without it(p<0.01).The Extra Tree Classifier model achieves the best performance on four metrics.It assesses Upper Cross Syndrome with accuracy of 82.4%(95% CI,75.7%-87.9%),sensitivity of 80.6%(95% CI,71.6%-87.7%),specificity of 85.5%(95% CI,74.2%-93.1%),positive predictive value of 90.2%(95% CI,82.2%-95.4%),negative predictive value of 72.6%(95% CI,60.9%-82.4%)and the area under the ROC curve for the model is 0.83.The results of the feature ablation experiments conducted on the Extra Tree Classifier show that the best combination of features is "angle1(a1),angle2(a2),angle3(a3),angle4(a4)",with a1 and a2 contributing more to the performance of the model and a3 and a4 contributing less to the performance of the model.Conclusions: Taller adolescents may be more likely to develop crossover syndrome;the Open Pose-based human key-point recognition algorithm can effectively extract features for assessing crossover syndrome in adolescents.