Machine learning using clinical variables to screen for depression and anxiety in people with early multiple sclerosis

Braxton Phillips, Sarah A. Morrow, Harasees Singh, Jiwon Oh, Shannon Kolind, Larry D. Lynd, Alexandre Prat, Roger Tam, Anthony Traboulsee, Scott B. Patten,
Machine learning using clinical variables to screen for depression and anxiety in people with early multiple sclerosis,
Multiple Sclerosis and Related Disorders,
Volume 104,
2025,
106761,
ISSN 2211-0348,
https://doi.org/10.1016/j.msard.2025.106761.
(https://www.sciencedirect.com/science/article/pii/S2211034825005036)
Abstract: Background
Depression and anxiety are more prevalent in people with multiple sclerosis (pwMS) than in the general population and are thus common psychiatric comorbidities in MS. As such, screening for psychiatric comorbidities in pwMS is an important component of MS care. However, large-scale screening efforts using scales that measure depressive symptoms with the goal of identifying a subgroup of pwMS with a high probability of depression, and thus in need of further assessment, have not been successful. Machine learning algorithms using routinely collected clinical information may be able to serve the same purpose without the encumbrance of scale administration and scoring.
Methods
We used baseline clinical and demographic data collected in the Canadian Prospective Cohort Study to Understand Progression in MS (CanProCo). The Patient Health Questionnaire-9 (PHQ-9) was used to screen for symptoms of depression and the Generalized Anxiety Disorder-7 (GAD-7) for symptoms of anxiety. Machine learning with elastic net was used to develop logistic regression models to predict participant scores above traditional cut-off scores ≥10 that are indicative of clinically significant depression and anxiety.
Results
Machine learning with elastic net produced a model that was able to predict scores ≥10 on the PHQ-9 in participants in the testing dataset with a high area under the curve (AUC = 0.927). Prediction of scores ≥10 on the GAD-7 in participants in the testing dataset was modest (AUC = 0.813). Final multivariable logistic regression models found that increased self-reported psychosocial fatigue (OR: 1.55, p < 0.001 and OR: 1.27, p = 0.0026), increased self-reported cognitive fatigue (OR: 1.08, p < 0.001 and OR: 1.08, p < 0.001), and number of comorbidities (OR: 1.26, p = 0.0015 and OR: 1.15, p = 0.041) were predictive of scoring above the cut-off on PHQ-9 and GAD-7, respectively.
Conclusion
The identification of psychiatric comorbidities such as depression in pwMS could be facilitated by making use of clinical variables with similar success to direct administration of rating scales.
Keywords: Depression; Anxiety; Machine learning; Screening