The SDS [42] was used as the primary tool to determine the depression levels of college students. It is a 20-item Likert-type instrument (4-point scale) to measure depressive affect and related symptomatology during a week by seeking responses to questions such as “I feel downhearted and blue,” “Morning is when I feel the best,” and “I have crying spells or feel like crying.” Participants were asked to choose one of the following responses to each of the questions: 1=rarely or none of the time (<1 day), 2=some or a little of the time (1-2 days), 3=occasionally or a moderate amount of the time (3-4 days), 4=most or all of the time (5-7 days).

SDS is one of the most common screening tests used by clinicians and psychiatrists. Participants’ raw scores ranged from 20 to 80. The higher the scores, the higher the occurrence of symptomatology. The standard point of the SDS is obtained by multiplying the raw score by 1.25. According to the Chinese norm for the general public, scores of 53 to 62 are classified as mild depression, 63 to 72 moderate depression, and 72 to 100 severe depression. SDS has an adequate item homogeneity; Knight et al [43] reported an alpha coefficient of .79 in a large community sample, and Gabreys and Peters [44] reported an alpha value of .88 in a sample of depressed patients.

In this paper, we developed an online questionnaire system and only allowed students connected to the campus network to participate in it. In order to control the accuracy of the questionnaire results, we recorded the duration of answering and removed students who completed the form in less than 5% of the overall time distribution and used the Beck Depression Inventory (BDI) [45] as a supplementary questionnaire to detect biased answers. As conclusions drawn by SDS and BDI are strongly related to each other theoretically, we deleted very contradictory items (eg, when the SDS showed that one student suffered from moderate to severe depression while the BDI indicated no depression). We also excluded participants who never had breakfast at the college canteens because they may be off-campus students.

We fully considered the privacy of students. First, we signed a data confidentiality agreement with the information management department and passed the ethical review of the University Department of Medical and Life Sciences (No 2018YXY24). Second, we indicated the purpose of the experiment, the type, and the time period of the use of the data on the first page of the questionnaire to let the participants know about this study.

With respect to the feature extraction, we designed a process to meet the regulations. We first gave the objectives and data requirements to the institutional review board and the information system administrator (ISA). After that, we got the metadata, such as field names with their meanings, data types, etc, of related tables from ISA. Subsequently, we provided detailed data requirements and the core code of data preprocessing. Furthermore, ISA wrote other auxiliary codes (eg, database accessing codes) and ran the whole preprocessing code on their own computers, and we finally got the features which did not contain any sensitive information such as students’ names and home addresses. In the whole process, we did not access any school information systems and therefore could not obtain any other private data. The process is illustrated in Figure 1.

The research conducted by Kessler et al [46,47] showed that depression episodes characterized as mild to severe have an average duration of 13.8 to 16.6 weeks, and very severe episodes have a mean duration of 23.1 weeks. The overall mean duration is 16 weeks. As one semester (18 weeks) is a complete cycle of student life on campus, we used data generated in the first semester of 2017 academic year (from September 11, 2017, to January 10, 2018) as the research subject for the study. The questionnaire data were collected from December 29, 2017, to January 28, 2018, and social relationship data were generated from November 20, 2017, to January 10, 2018.

In order to determine the symptom groups contained in the SDS, Zung divided the SDS items into three groups based on the overt content of symptoms: pervasive affect (depressive affect and crying spells), physical equivalents (diurnal variation, sleep disturbance, decreased appetite, decreased libido, weight loss, constipation, tachycardia, and fatigue), and psychological equivalents (confusion, psychomotor retardation and agitation, hopelessness, irritability, indecisiveness, personal devaluation, emptiness, suicidal ideation, and dissatisfaction) [42]. Other literature [40,48,49] adopted factor analysis to identify common clusters across different types of people, but this is difficult because the profiles of depressive symptoms vary in different populations [48].

Although some literature proposed factor structures of college students, their conclusions may not be suitable for Chinese students because of the differences in the populations. Therefore, we redivided the SDS questionnaire into different depression symptom groups using factor analysis. The accumulated score of items that belonged to each symptom group was used to represent the severity of the symptom.

Out of 502 undergraduate students who participated in the research, 466 students were selected as research subjects and their detailed information is shown in Table 1.

Of the participants, 69.5% (324/466) were boys and 30.5% (142/466) were girls. Ages ranged from 17 to 23 years. Grades ranged from one to six and the majority were first-year students (214/466, 45.9%). The reason there were two six-grade students in the sample is that some students repeated a grade because of various reasons (eg, receiving medical treatment).

The mean depression score of male students was 46.38, while the females’ was 49.14. The ratio of students suffering from depression was 25.32%, with mild depression (16.31%) taking the majority, and a small proportion of students were found suffering from moderate (6.44%) to severe (2.58%) depression. We used the chi-square test to determine the gender difference in depression rate, and the results are shown in Table 2.

The Mann-Whitney U test is a widely used nonparametric test to decide whether two groups of samples are derived from the same population. Since most of the features in this study deviated from the normal distribution, the nonparametric Mann-Whitney U test was used to determine whether there were differences between the two groups with respect to the behavior on campus.

The test indicated there were 25 features significantly different between depressive and nondepressive groups, and the results are listed in Table 3.

The result showed χ2=10.0 (P=.02), which indicated different genders had different rates of depression levels. Post hoc testing demonstrated that female students had a higher depression rate than male students (odds ratio [OR] 0.54, 95% CI 0.349-0.835); more specifically, female students had a higher risk of mild depression (OR 0.473, 95% CI 0.286-0.783), while there was no statistical difference in moderate and severe depression.

Eating behavior was found to be different between the two groups of students: depressive students (u=45.81) skipped breakfasts significantly more often than nondepressive (u=52.77) students (z=–2.25, P=.03). The same was observed for lunch and dinner. In addition, the breakfast time of depressive students was later than that of nondepressive students (z=–2.91, P=.004), and lunch and dinner time were also later.

With regard to daily schedule preferences, the number of depressive students entering the dorm in the morning was significantly less than that of nondepressed students. In addition, the night activity of depressed students (ie, the average number entering or leaving the dorm between 8 pm and midnight was significantly busier than that of nondepressive students.

Regarding social interaction, all 5 social features of depressive students were significantly lower than their counterparts. Figure 4 illustrates the distribution of social behavior features of the two groups. The box plot indicates that the lower quartile, average, median, upper quartile and maximum of the depressive group are all lower than nondepressed students. It is, in turn, revealed that the student card–based social relationship features proposed in this paper can reflect students’ social characteristics.

The number of poverty subsidy applications of depressive and nondepressive students was statistically different (χ2=4.2, P=.04), with the application rate of poverty subsidy in the nondepressed group 4.90% and depressed group 10.17%, suggesting that there is a higher rate of depression among economically disadvantaged students.

Correlation analysis showed that there was a high correlation between the vast majority of items on the SDS, which means it is possible to use a few factors to summarize most of the information contained in the original 20 entries. Principal component factoring with promax (oblique) rotation was adopted to evaluate the factor structure of the SDS in this study. The Bartlett test of sphericity was significant (χ2=2505.6, P<.001), and the Kaiser-Meyer-Olkin measure of sampling adequacy was 0.92, indicating the rationality of principal component factoring.

The eigenvalues of the first 4 principal components were greater than 1, and the variations were 30.15%, 8.01%, 5.72%, and 5.34%, respectively. The cumulative variance was 49.23%. The rotated factor pattern matrix is given in Table 4.

The first factor (F1) was composed of 10 items: decreased libido, personal devaluation, psychomotor retardation, dissatisfaction, decreased appetite, confusion, hopelessness, emptiness, indecisiveness, suicidal ideation. The second factor (F2) was composed of 6 items: crying spells, irritability, fatigue, sleep disturbance, psychomotor agitation, and depressed affect. Factor III (F3) consisted of 3 items: constipation, weight loss, tachycardia. Factor IV (F4) consisted of only one item: worse in the morning. Unlike factor analysis results of some literature [40,48,49], diurnal variation (worse in the morning) was the only item that has large loading in F4. We summed up the scores of all items belonging to each factor or symptom group as the score or severity of the symptom.

Table 5 shows the Spearman correlation between daily routine features and the four factors and the total SDS score.

Among the 4 factors, F4 can be reflected in daily routine most, and the reason for the correlation may be that depression disrupted circadian rhythms [50]. Students with higher levels of depressed mood (F2) and somatic (F3, F4) depressed symptoms partially result in increased night activity, while a decrease in morning activity is partially caused by cognitive (F1), depressive mood (F2), and somatic (F4) symptoms (see Multimedia Appendix 1 for detailed information).

Depression symptoms can result in changes in appetite [29]. Intuitively, these changes may be reflected in the dining data. To test this hypothesis, we performed the Spearman rank correlation analysis between meal habits and depression symptoms, and the test results are shown in Table 6.

As can be seen from the table, F4 correlates with most meal habit features with high confidence. Feeling worse in the morning can lead to poor meal habits, such as a high rate of skipping meals and irregular eating.

As most students should accomplish the majority of courses in the first two years of college life, first-year students are empirically found to be more serious about their study compared with higher year students, therefore we only use first-year student data (214 students) to investigate associations between academic performance and the severity of depression. The result of Spearman correlation is reported in Table 7.

Compared with other factors, F1 has the greatest impact on performance, especially for the two extremes, fail and excellent. When the F1 score is high, it is easier to fail exams or to get general grades. However, the total SDS score was only found to significantly correlate with the fail course rate and the number of failed courses. The table suggests that depression makes it easier for students to have bad academic performance, even if the total score does not reach the threshold of depression.

The distribution of social behavior features is normally distributed, so we tested the Pearson correlation between features and 4 facts and SDS score. The result is shown in Table 8.

The table indicates that F4 had the largest effect on social activity, students who felt bad in the morning were more reluctant to share meal times with their friends, while the total SDS score only has a weak relationship with social activity.

The relationship between the frequency of library access and depression was also explored, F4 was found to be negatively correlated with the number of library visits (r=–.179, P<.001). No difference was found in the acquisition of scholarships.

Being a serious mental disorder, depression affects the life of patients in both mental and physical aspects. There is a critical need to identify behavior characteristics for better identifying and even helping the treatment of depression; however, to the best of our knowledge, research on depression among college students with campus data has not been conducted.

The meal habits of depressive students in this research were different from that of nondepressive students; this finding is congruent with the finding of Lee et al [29] that implied depression may affect the dietary habits. Our experiments indicated students with depression tended to have bad eating habits. Factor analyses indicated feeling bad in the morning was a contributor to the phenomena; furthermore, irregular breakfast time was caused by complicated symptoms except for cognitive factor.

Depressed students were found to have fewer morning activities and more night activities; furthermore, factor analyses indicated that decreased morning activity was affected by factor 4 (feeling bad in the morning), while increased evening activity was partially caused by factor 2 (depressed mood) compared with other symptoms.

Our findings on daily routine and meal habits are consistent with the well-established literature that suggests biologic rhythm is associated with depression [35-38,51-53]. Disruptions in behavioral patterns caused by depression during waking hours includes not only the volume of activity but also the patterns of behavior [53,54]. These pattern changes may be a result of the genetic and hormonal factors [36] implicated in depression-related circadian rhythm changes [53,54].

With respect to social interaction pattern, the Mann-Whitney U test and correlation analyses demonstrated that depression impaired social activity, and this conclusion is congruent with previous studies [38,55]. Furthermore, our study demonstrated that social activity extracted in this research was mainly affected by diurnal variation (F4), while cognitive symptoms (F1) and somatic symptoms (F3) had no connection with decreased social activity.

Many studies have investigated the negative impact of depression on academic performance [5,26,56], and our study goes further on this topic. Factor analyses demonstrated that there was a negative correlation between depression and course fail rate among first-year students. Moreover, by investigating the relationship between academic performance and each factor, we found that the cognitive factor (F1) was the most relevant factor for poor academic performance. Students with high cognitive symptom were more likely to fail exams and less likely to achieve excellent academic performance [57].

Finally, through the chi-square test we found that the proportion of poverty subsidy application was not the same between depressive and nondepressive students. The application proportion of the depressive group was significantly higher than the nondepressive group, which indicates that family economic status is one of the impact factors in depression. This conclusion fits into previous studies, which demonstrated poor socioeconomic background is associated with depressive symptoms within each country [5,56].

Although our research reveals that some behavioral characteristics of depressive students can be captured by campus data, the results are preliminary and a number of limitations must be mentioned.

First, the work has not covered all aspects of students’ life on campus (eg, data related to location, physical exercise, after school clubs, course selection) due to the limitation of school information infrastructure. With respect to meal habits, because there are other places for students to eat except for the campus canteen, it is impossible to accurately obtain all the dining information. Thus, the term meal habits described in the paper only represents the meal habits in the campus canteen.

Second, although some patterns about depressive students’ daily routines were found, specific reasons need to be further clarified through a questionnaire survey.

Third, this research adopted the method of self-reported questionnaire survey to determine whether participants were depressed, which may bring about the pervasive problem of response bias [58].

Finally, students with different levels of depression may behave differently. Because of the low number of students who suffered from moderate to severe depression, behavioral differences among mild, moderate, and severe depression were not explored.

This paper proves that campus data can reflect student depression status and indicates that depression may be predicted by using machine learning techniques. Because timely rescue of potentially depressed students can improve their academic performance and reduce their own pain and suicide risk, depression prediction has good practical effects.

Privacy presents a major problem. To precisely predict students’ mental health problems, too much privacy data such as online behavior, daily routine, family background, and consumption level may be involved. Thus, it is recommended that data providers convert or add noise to the original data to protect privacy. For example, provide researchers with differential privacy [59] processed data.

In addition, data-driven mental health prediction methods can lead to some negative effects, and the feedback loop is a topic worthy of consideration because it creates some problems.

The principle of educational equity may be violated. Managers may prefer not to grant scholarships to depressed students because they may perform poorly, and this would reduce their chances of participating in academic, community, and competition activities. This could be addressed by developing a mental health prediction system that can only be used by those whose jobs are related to mental health care. When the system finds a depressed student, a psychological counselor will contact the student directly to decide whether it is appropriate to recommend that they ask for psychological counseling services according to the actual situation. Because of direct communication between the counselor and depressed student, no other students or teachers would know the specific situation, therefore eliminating the problem of educational inequality that may be caused by depression.

Although data management rights belong to the information technology department, it does not mean that the department has full rights to the data. Some operations on the data are forbidden to perform unless permitted by students. Some students may be concerned about mental health issues or personal privacy being exposed, which may result in a refusal of authorization. Therefore, some depressive students will not be detected by the system. To address this, we will popularize mental health knowledge on campus to eliminate student stigma about mental illness and explain to students how the system works to eliminate their concerns about privacy.

The campus information systems run all the time and record all aspects of students’ living on campus. Compared with the traditional questionnaire survey method, these data can reflect behavior and psychological activities more objectively and realistically, and there is almost no cost to obtain this data compared with other types of data (eg, mobile phone use) [54]. This paper proves that depressive students have different behavioral characteristics than nondepressive students by using campus data and the symptoms of depression can also be reflected in the data, supporting scholars’ behavioral models of depression [24]. This paper also provides new ideas for discovering human behavior patterns associated with depression and other mental health disorders. If combined with other types of campus data (eg, network connection event logs [60]), it is possible to achieve an artificial intelligence–based student depression prediction method.