Speech Features Models In Suicide Risk Recognition In Unipolar And Bipolar Depressive Patients Based On Machine Learning

Background and PurposeSuicide is a major global public health problem.Depressive disorder is highly related to suicide.Early recognition is of great significance to suicide prevention.However,the current screening methods have obvious limitations,with low recognition accuracy and subjective influence.Finding objective and effective recognition methods has become a research trend.Voice has the advantages of fast and convenient collection,less restriction,low cost,and no intrusion into the human body,which had attracted the attention of many researchers.Therefore,the purpose of this study is to explore the effectiveness of speech feature models to identify suicide risk in patients with unipolar and bipolar depression based on machine learning.MethodsParticipants were recruited from universities and psychiatric clinic.Voice data were collected by using a design of 2 ×(group:healthy control vs.case)× 3(task format:interview vs.reading aloud vs.picture description)× 3(emotional valence:negative vs.neutral vs.positive).The collected data were performed preprocessing and extracted speech features which used widely in this research field Support Vector Machine(SVM),Ensemble Learning(EL),Naive Bayesian(NB)and k-Nearest Neighbor(k-NN)were used to model and analyze.Accuracy(ACC),Specificity(SPE),Sensitivity(SEN)and Area under Curve of Receiver Operating Characteristic(AUC)were used for model evaluation.Results(1)The optimal model of speech recognition unipolar and bipolar depression was constructed by EL(ACC=88.89%,SPE=91.84%,SEN=85.08%,AUC=0.885),average ACC of model was 74.37%,and average AUC was 0.727.For female group,the optimal model was constructed by EL(ACC=88.97%,SPE=93.75%,SEN=83.08%,AUC=0.884).The average ACC and AUC of models were 75.58%,0.748.While for male group,the optimal model was constructed by SVM(ACC=92.08%,SPE=92.86%,SEN=91.00%,AUC=0.919).The average ACC and AUC of models were 77.60%and 0.766.(2)For the task format question and answer had the best effects,which constructed by EL(ACC=92.76%,SPE=95.91%,SEN=88.72%,AUC=0.923).The average ACC and AUC of models were 77.63%and 0.766.For emotional valence,positive emotions were the best,which constructed by SVM(ACC=87.62%,SPE=91.23%,SEN-82.95%,AUC=0.871).The average ACC and AUC were 71.16%and 0.700).(3)The optimal model of speech recognition for suicide risk in patients with unipolar and bipolar depression was constructed by EL(ACC=91.86%,SPE=96.22%,SEN=84.55%,AUC=0.904).The average ACC and AUC were 77.37%and 0.744.From gender perspective,the generalization performance of male group was still better than that of female(average ACC:80.75%vs.77.82%,average AUC:0.801 vs.0.747).(4)The accuracy of speech recognition patients’ suicide risk was affected by task form and emotional valence.Relatively speaking,interview speech(constructed by EL,ACC=93.23%,SPE=98.80%,SEN=84.00%,AUC=0.914)and negative emotion valence(constructed by SVM,ACC=88.76%,SPE=96.43%,SEN=75.76%,AUC=0.861)had the best generalization performance.(5)The prediction effects of the multi-index model were better than that of the single speech feature model(average ACC:80.85%～84.07%vs.77.37%,average AUC:0.783-0.822vs.0.744),and the optimal model was still constructed by EL.Conclusion(1)Speech could be used as objective indicator to identify the suicide risk in patients with unipolar and bipolar depression.(2)Different task forms and emotion valence influenced the model predictive performance and interview speech had the best predictive performance.Positive valence was best when identifying unipolar and bipolar depression,while negative valence was best when identifying suicide risk.(3)The combined subjective and objective multi-index models were better than the single-index model.It was suggesting that we should consider from multiple angles to give full play to the role of speech in future screening work.(4)Among the four machine learning algorithms,ensemble learning had excellent performance and it can be given priority.Innovation(1)Considering the limitations of current screening methods,we explored the efectiveness of speech indicators.It has important reference value for screening work and has broad application prospects in the future.(2)Our study enriched the materials of speech-related research,especially suicide-related filed and paved the way for further research work.At the same time,it extended the thinking on suicide prevention.