Identification of various user group in university counseling scene using machine learning algorithm

Abstract

There are increasing need for improving university students' mental health. Still, the quantitative and qualitative abilities of the university counseling center are insufficient to handle the increasing need and severity. In addition, it is difficult for individual counseling centers to solve such problems due to lack of numbers in counselors and low budget. Therefore, to solve current problems by increasing the service efficiency of the university counseling center service, identification of various user groups was done using machine learning algorithm based on initial stage data of counseling service. To be specific, user service, additional clinical and latent group was identified to reduce counselor’s burden at initial stage of service and provide reference for clinical decision and future service planning. This study utilized data acquired from UNIST healthcare center and analyzed in two major steps. First, service (counseling or clinical treatment with drug use) and clinical (suicidal-risk and potential dropout) group classification was done using supervised learning algorithms and identified important feature in classifying each group. Then, the latent groups reflecting the detailed characteristics of users was analyzed by using latent class analysis. Laten user group identification detected 5 different latent groups (lower risk, lower/moderate risk, moderate risk, higher risk with sleep issue, sleep problem group) and their distinctive characteristics. Current study successfully detected meaningful reference for university counseling service using data focused on the initial stage of the service. Analyzing service effectiveness and using it as reference for converting counseling service into clinical treatment according to current study results will increase overall service effectiveness provided to individual users. The result of suicidal-risk group classification identified similar features from prior researches without using additional screening tools. Interestingly, dropout group classification results identified features that were not found in prior research which can be used in future service planning to prevent user dropout during the service. After the classifications of various user groups were conducted, improving the result of ensemble modeling using stacking classifier was done to achieve higher performance in type 2 error of classification results. Latent group identification found sub-groups that can be applicable to existing counseling services and possible customized clinical approaches can be provided to individual latent groups. Further studies including 1) improving machine learning algorithm performance by developing data collection methods that reflect user characteristics 2) better, service effectivity analysis using overall service records and 3) applying studied researches to other university counseling centers will also contribute to reducing individual counselor’s burden and enhancing university counseling center service effectiveness.