JMIR Publications

ABSTRACT

Background:

Anhedonia and depressed mood are considered the cardinal symptoms of major depressive disorder. These are the first two items of the Patient Health Questionnaire-9 (PHQ-9) and make up the ultra-brief PHQ-2 questionnaire used for pre-screening depressive symptomatology. The pre-screening performance of alternative PHQ-9 item pairings is rarely compared with that of the PHQ-2.

Objective:

Use a data-driven Machine Learning (ML) approach on the PHQ-9 items to find and validate the most predictive two-item depressive symptomatology ultra-brief questionnaire. Test the generalisability of the best pairings found on the primary dataset, with six external datasets from different populations to validate their use as pre-screening instruments.

Methods:

All thirty-six possible PHQ-9 item pairings (each yielding scores of 0-6) were investigated using Machine Learning-based methods with logistic regression models. Their performances were evaluated on the classification of depressive symptomatology, defined as PHQ-9 scores ≥10. This gave each pairing equal opportunity and avoided any bias in item pairing selection.

Results:

The ML-based phq2&4, the depressed mood and low energy item pairing, and phq2&8, the depressed mood and psychomotor retardation or agitation item pairing, were found to be the best on the primary dataset training split. They generalised well on the primary dataset test split with Area Under the Receiver Operating Characteristic Curves (AUC) of 0.954 and 0.946, respectively, compared with an AUC of 0.942 for the PHQ-2. The phq2&4 had a higher AUC than the PHQ-2 on all six external datasets and the phq2&8 was higher than the PHQ-2 on three. The logistic regression probability thresholds that maximised Youden’s index (an unweighted average of sensitivity and specificity) during cross-validation on the primary dataset were applied to the phq2&4 and phq2&8 models. The phq2&4 had the highest Youden’s index on two external datasets and the phq2&8 was highest on another two. The PHQ-2 ≥2 also had the highest Youden’s index on two external datasets, joint highest with the phq2&4 on one, but its performance fluctuated the most. The PHQ-2 ≥3 had the highest Youden’s index on one external dataset. The sensitivity and specificity achieved by the phq2&4 and phq2&8 were more evenly balanced than the PHQ-2 ≥2 and ≥3 cutpoints. For the PHQ-2, the former cutpoint excessively favoured sensitivity over specificity and vice versa for the latter.

Conclusions:

The PHQ-2 was not a superior pre-screening instrument to other PHQ-9 item pairings. Evaluating all item pairings showed that, compared with alternative partner items, the anhedonia item underperformed alongside the depressed mood item. This suggests the inclusion of the anhedonia item in ultra-brief questionnaires may be a relatively arbitrary choice. The use of the PHQ-2 to pre-screen for depressive symptomatology could result in a greater number of misclassifications than alternatives such as the phq2&4 and phq2&8.