Alternative scoring of the Patient Health Questionnaire-9 in neurological populations: an approach based on a predictive algorithm deriving from individual item scores

Abstract

Objective: The study objective was to assess whether machine learning methods could improve predictive performance of the PHQ-9 for depression in patients with neurological disease. Specifically, we assessed whether a predictive algorithm deriving from all nine items could outperform the tradition of summing the items and applying a cut-point.

Method: Data from the NEEDS Study was used (n = 825). Demographic data, PHQ-9 scores, and MDD diagnoses (via the SCID) were obtained. Logistic LASSO, logistic regression, and non-parametric ROC analyses were performed. The ROC curve was used to identify the optimal cut-point for regression-derived predictive algorithms using the Youden method.

Results: The traditional approach to PHQ-9 scoring had a classification accuracy of 85.1% (sensitivity: 84.5%; specificity: 85.2%). The logistic LASSO regression model had a classification accuracy of 85.6% (sensitivity: 83.3%; specificity: 86.1%). The logistic regression model had a classification accuracy of 85.8% (sensitivity: 91.4%; specificity: 84.8%). Both models had similar areas under the curve values (logistic LASSO: 0.9097; logistic regression: 0.9026).

Conclusions: The current cut-off threshold approach to PHQ-9 scoring and interpretation remains clinically appropriate.

Keywords: Major depressive disorder; Neuropsychiatry; PHQ-9.