Use of the Taiwanese Depression Questionnaire and the AD8 Questionnaire for Screening Depression in Older People in Communities

Background Depression presents with emotional and somatic symptoms, and sometimes cognitive complaints. On the other hand, depression is one of the important psychiatric symptoms of mild cognitive impairment and dementia. Methods Residents who were older than 50 years reported their depressive tendency using the Taiwanese Depression Questionnaire (TDQ) and cognitive complaints using the AD8 questionnaire and were assessed using the Mini-International Neuropsychiatric Interview (MINI) and for objective cognitive evaluation using the Mini-Mental State Examination (MMSE). Results TDQ score (OR 1.154, p = 0.003) and AD8 score (OR 1.769, p = 0.018) were statistically signicant in predicting current major depressive disorder (MDD) when adjustment were made for age, sex, sleep quality and cognitive performance. However, in elderly people with age ≥ 65, TDQ score failed to distinguish a diagnosis of current MDD from no such diagnosis (AUC 0.780, p = 0.063). A linear combination of TDQ and AD8 scores provided a new predictive score that successfully separated elderly people with MDD from those without it (AUC 0.875, p = 0.013). Finally, decision-tree learning was used to generate a classication tree for classifying people with and without current MDD. At the rst decision level, value less than 29 from the sum of TDQ score and 4 folds of AD8 score 100% excluded people without current MDD. Conclusions The self-reported response to the TDQ is a feasible way to identify MDD in community-dwelling people. Combining TDQ and AD8 scores further improved detecting depression in older people in communities. coecient − 0.688, p < 0.001) and mental health, per the SF-36 questionnaire (correlation coecient − 0.603, p < 0.001). These results suggest that sleeping quality and quality of life worsen as depressive traits increase.

prediction model, cut-off score was determined by the highest Youden's J statistics [(sensitivity + speci city) -1] in the model. Since subjective cognitive complaints usually increase with age, the dataset was separated into a subset for age ≥ 65 and a subset for ages 50-65 and ROC curves were obtained for these two groups. Finally, decision tree learning was used to generate a classi cation tree from the observed demographic data, TDQ score, and AD8 score to concluding MINI-MDD as the target value. Statistical analysis was performed using IBM SPSS Statistics 20 software.

Results
Of 127 community-dwelling residents from the northern coastal area of Taiwan, nine were excluded as subjects for incompletely answering the TDQ, AD8 or MINI. A total of 118 residents were thus enrolled. Based on MINI diagnoses, of those with current MDD 33.3% were widowed and 22.2% were divorced, whereas only 10.7% and 6.8% of those without current MDD (p < 0.05) were widowed and divorced, respectively. Both TDQ scores (25.56 ± 10.90 versus 11.05 ± 9.79, p < 0.001) and AD8 scores (4.00 ± 2.45 versus 1.94 ± 2.09, p < 0.01) were higher in the group with current MDD. A higher percentage risk of suicide was identi ed in people with current MDD (55.6% versus 0%). The group comparison revealed no difference in age, sex, years of education, income level or cognition status in terms of MMSE score. (Table 1)  Spearman's rho statistics were used to evaluate correlations among TDQ, age, years of education, income level, AD8 score, MMSE score, PSQI score and SF-36 score. Statistics revealed a weak correlation between TDQ and AD8 scores with a correlation coe cient of 0.386 at a signi cance of p < 0.01 (Table 2). Depressive tendency slightly increased with subjective cognitive complaints. The depressive score on the TDQ was also moderately correlated with sleep quality in terms of PSQI (correlation coe cient 0.561, p < 0.001), quality of life with respect to general health (correlation coe cient − 0.564, p < 0.001), vitality (correlation coe cient − 0.688, p < 0.001) and mental health, per the SF-36 questionnaire (correlation coe cient − 0.603, p < 0.001). These results suggest that sleeping quality and quality of life worsen as depressive traits increase. In the logistic regression model for the MINI diagnosis of current MDD, AD8 and TDQ scores were both included as variables. TDQ score alone and AD8 score alone had 1.127 (p = 0.001) and 1.463 (p = 0.012) folds of odds of indicating current depression in MINI respectively (Table 3, models 1 and 2). Accounting for whether someone is elderly (age ≥ 65) and their sex, TDQ score signi cantly in uenced the diagnosis of depression (Table 3, model 3). When sleep quality in terms of PSQI and objective cognitive assessment in terms of MMSE are considered along with old age and sex, both TDQ and AD8 scores signi cantly affected the likelihood of a depressive disorder (odds ratios OR 1.154, p = 0.003 and OR 1.769, p = 0.018, Table 3, model 4). The ROC curve was used to evaluate the performance of self-reported questionnaires in predicting a diagnosis of major depression. The AUC was 0.835 for the TDQ score (p = 0.001) and 0.751 for the AD8 score (p = 0.013). An AUC over 0.5 was regarded as indicating that target values could be effectively classi ed. When using the linear combination of TDQ and AD8, the sum of TDQ score and four times the AD8 score predicted a MINI diagnosis of major depression with an AUC of 0.887 (p < 0.001) (Fig. 3-A). Calculating Youden's J statistics as [(sensitivity + speci city) -1] yielded the best cutoffs of TDQ for predicting MDD, which were identical to that originally proposed for clinical use: a TDQ score of 19 with a sensitivity 0.889, a speci city of 0.780, a positive predictive value of 88.9% and a negative predictive value of 77.3%. A linearly combined TDQ and AD8 score of ≥ 31 was feasibly predicted MDD diagnosis by MINI for community-dwelling people, with a sensitivity of 1.000, a speci city of 0.789, a positive predictive value of 100% and a negative predictive value of 77.1%.
Owing to the possibility of atypical presentations, such as subjective cognitive complaints in elderly people with major depression, enrollees were separated into groups of age ≥ 65 and age 50-65, to examine whether age in uenced likelihood of major depression. In the elderly group (Fig, 3-B), prediction by TDQ score or AD8 score alone was unsatisfactory, yielding AUCs of only 0.780 (p = 0.063) and 0.627 (p = 0.398) respectively, and failing to reach statistical signi cance. The sum of the TDQ score and four times the AD8 score, however, offered much improved predictive performance, with an AUC of 0.875 (p = 0.013). In the group with age 50-65 ( Fig. 1-C), the TDQ score, AD8 score and their linear combination were less predictive, but the linear combination remained the strongest predictor; the AUC values were 0.892 (p = 0.004), 0.804 (p = 0.026), and 0.918 (p = 0.002) respectively.
To classify people with and without major depression, a basic machine learning algorithm with decision tree learning was used to analyze the effects of such variables as age, sex, education year, level of annual income, marital status, TDQ score, AD8 score, and the sum of TDQ and four times the AD8 score. The model thus obtained had two levels and four nodes. The linear combination of TDQ score + 4*AD8 score was at level 1, and a value of less than 29 at node 1 indicated no current MDD while a value of over 29 at node 2 indicated current MDD. The second level came from node 2 by sex; the female gender at node 3 favored current MDD whereas the male gender favored no current MDD (Fig. 2).

Discussion
This study is a cross-sectional study that was carried out by the Community Medicine Research Center of Keelung Chang Gung Memorial Hospital, which ran a cohort study in northern costal Taiwan. Using self-reported responses to questionnaires on depression (TDQ) and subjective cognitive complaints (AD8), this work seeks to screen people with major depression. At total of 118 community residents were recruited after comprehensive psychiatric assessment using MINI identi ed nine people with current MDD. According to group analysis, correlation analysis, logistic regression modeling, AUC of ROC curves, and a decision tree model, TDQ score was useful in distinguishing people with current MDD from those without. Based on the ROC curve and decision tree models, a new score equal to a linear combination of the TDQ and AD8 scores outperformed TDQ score alone.
The limitations of the study included the small number of people with current major depression, which might have led to biased statistical results. Second, the recruitment of community-dwelling people and their visits to hospital for complete psychiatric evaluation required their cooperation. Those willing to undergo that evaluation might have been more active and had a greater ability to move than the others. People with major depression might have had less motivation and therefore have been more likely to refuse to participate. Consequently, the number of people with MDD could be under-estimated. Finally, grouping the enrolled people by age group to generate ROC curves further reduced the statistical power of the determinations made based on those groups.
Depression is a multifactorial disease that links to socioeconomic function, culture and interpersonal relationships. Comparisons of patients with and without current MDD demonstrated that their percentages married differed (Table 1). Marital disruption, including divorce and widowhood, precipitate depression [21,22] and this fact may explain the high percentage of non-married people with current MDD. Other objective criteria like demographic characteristics, education, and nancial status did not precipitate major depression. However, in the decision tree model, under the node of score summation of TDQ + 4*AD8 (Fig. 2), female gender favored the identi cation of depression in people with a score greater than 29 points. Although the between-group comparison failed to reveal a signi cant difference in sex ratio (p = 0.150, Table 1), people with MDD were more likely to be female (eight of nine) than those without MDD. The statistical insigni cance of the chi-square statistics might be explained by small number of cases in the MDD group; the clinical signi cance of female gender in predicting MDD at the decision tree is logical because depression is 1.7 times as prevalent in females than males (5.5 to 3.2%) [23].
Depression is commonly concomitant with cognitive complaints, as determined in our earlier study of subjective cognitive decline [24]. TDQ is a self-reporting depressive questionnaire on somatic and emotional aspects. AD8 is a self-reporting questionnaire that covers various cognitive domains in daily life. Correlation tests in Table 2 revealed a correlation between TDQ and AD8 scores but not a strong one. Therefore, depression complaints and subjective cognitive complaints should be considered together, and collinearity was not a problem in the regression model. In logistic regression (Table 3), model 4 reveals that TDQ and AD8 scores both predicted MINI-current MDD; according to the ROC curve in Fig. 1-A, AUC was improved by combining TDQ and AD8 scores in a new predictive score. Therefore, co-considering depressive and cognitive complaints is a feasible means of improving the identi cation of major depression.
The complex relationship between cognitive impairment and depressive disorder has a long history; debates have addressed prodromal depression in later dementia [25], depression as a risk factor of dementia [26], depression's mimicking dementia, called pseudodementia [27], and depression as a psychiatric symptom of various types of dementia [28,29]. One of the most important indicators of cognitive decline is a self-sensation of cognition dysfunction [30]. In a systematic review, subjective cognitive impairment and depressive presentations were found to have reciprocal effects, and were frequently accompanied by anxiety and stress [31]. The severity of depression has been correlated with the degree of cognitive complaints [32]. Among various factors that mediate cognitive symptoms in depression, such as duration of current depression episode and presence of disability [33], age is important. Late-life depression (age ≥ 60) presents more memory impairment, worse verbal learning, and slower motor speed than depression in younger individuals [34]. Notably, atypical presentations of depression in the elderly were considered in the development of the Geriatric Depression Scale (GDS). In 1982, Yesavage noted that psychomotor retardation and passive refusal lead to mistaken identi cation of depression as dementia. Depression in the elderly is usually accompanied by experiences of memory or cognitive decline, and somatic symptoms of depression are commonly confused with other age-related somatic symptoms. The high frequency of cognitive complaints in cases of geriatric depression may be misleading but also useful in identifying depression in the elderly [35]. Since the features of depression in the elderly are distinct, geriatric depression represents a sub eld of affective disorders [8].
In this work, integrating subjective cognitive survey with a traditional depression survey showed provided additive value in identifying depression in elderly people aged over 65 years. Neither TDQ nor AD8 score alone su ced to identify depression in elderly, but a linear combination of AD8 and TDQ scores as a new predictor signi cantly improved the ROC curve ( Fig. 1-B). The TDQ was originally developed to suit Taiwan's cultural traits and forms of emotional expression [3], but did not consider the effect of age on presentations of depression. Given the special characteristics of geriatric depression, cognitive complaints, AD8 and the culture-speci c questionnaire TDQ were used in screening for depression; doing so improved the AUC in ROC curves; and the linear combination of TDQ and AD8 was critical in implementing the classi cation decision tree. A decision tree is basic machine learning model and is suitable for solving multifactorial diagnostic problems with hierarchical variables [36,37]. In the model herein, multiple variables were considered, but the linear combination TDQ + 4*AD8 was the best for identifying depression in community-dwelling older people. The cut-off value of TDQ + 4*AD8 was 29/30 in the decision tree model and 30/31 in the ROC curve based on Youden's J statistics; the closeness of these cutoffs indicate the consistency of this variable in predicting depression.

Conclusions
The TDQ can feasibly be used for screening for major depression in community-dwelling people. Incorporating subjective cognitive complaints using the AD8 questionnaire further improved the identi cation of depression in the elderly. A decision tree model that comprehensively weighted in uencing factors was developed to identify depression. Availability of data and materials: The datasets used and/or analyzed during the current study are available from the corresponding author on reasonable request.