JMIR Publications

ABSTRACT

Background:

One of the reasons why students come to counseling is when they are called on based on a self-report health survey result. However, there is no concordant standard for the call.

Objective:

This study developed a machine learning model to predict students’ mental health problems in one year and the following year using the health survey's content and answering time (response time, response timestamp, and answer date).

Methods:

The analysis was based on the responses of 3,561 undergraduate students from University A in Japan (a national university) who completed the health survey in 2020 and 2021 (response rate of 62.58%). With- and without- answering time conditions were compared. From 2020 data, a mental health problem in 2020 was predicted in analysis 1, while a mental health problem in 2021 was predicted in analysis 2.

Results:

Both analyses and conditions achieved adequate accuracy (for example, AUC-ROC: analysis 1 with answering time 0.838; analysis 1 without 0.825, analysis 2 with 0.760; analysis 2 without 0.765). In both analyses and in both conditions, the response to the questions about campus life (for example, anxiety, future, and schoolwork) had the highest impact (Gain 0.062-0.097, SHapley Additive exPlanations 0.031-0.042). SHapley Additive exPlanations of four to six feature variables from questions about campus life were included in the top seven. With answering time condition outperforms the without answering time condition in dichotomous measures (for example, precision: analysis 1 with answering time 0.76; analysis 1 without 0.57, analysis2 with 0.76; analysis 2 without 0.66), whereas the difference was unclear in probabilistic measures (for example, AUC-PR: analysis 1 with answering time 0.25; analysis 1 without 0.23, analysis2 with 0.21; analysis 2 without 0.23).

Conclusions:

These results show the possibility of predicting mental health across years using health survey data.