Predicting suicide risks using Machine Learning among medical students in Bangladesh

Abstract

Medical students face multiple challenges that can significantly impact their mental health. The intense academic pressure, exposure to death and long working hours make them vulnerable to stress, anxiety and sometimes depression. In Low and middle-income countries (LMIC), additional factors like social stigma, peer pressure and family issues like parental expectation can further contribute to their mental instability, ultimately leading to the ideation of suicide. It is found that in Bangladesh alone, the prevalence of suicidal ideation was 27.4% which is higher than the global prevalence. Even though the rates keep increasing, early signs of suicidal ideation go undiagnosed due to various reasons such as social and cultural stigma. Thus, this thesis brings forth a model developed for predicting suicidal risks among medical students in Bangladesh due to the complexity of this topic. It aims to contribute to the knowledge about mental health by creating a feasible machine learning model to identify suicidal thoughts and access to early intervention among Medical Students in Bangladesh. Three publicly available datasets were selected where multiple types of data such as symptoms, demographic & scale data were utilized to extract information through surveys. Four machine learning algorithms were employed on each dataset, after careful cleaning and preprocessing for all the models. For feature extraction, LASSO, Pearson’s Correlation Coefficient (PCC), Sequential Forward Selection (SFS), and Weighted Sum Feature Selection (WSFS) were utilized. When Random Forest was applied to the features selected by WSFS on Dataset A, it proved to be the best model with an accuracy of 0.97 and an AUC of 0.99, further supported by Cohen’s Kappa value of 0.95. LASSO selected features from all datasets performed well, and algorithms gave good results as well. Overall, the Random Forest performed better than Decision Trees, while Logistic Regression and SVM gave almost similar results. The combined dataset model, however, was not able to showcase good performance.