|
Abstract
|
Background
Screening for major depressive disorder (MDD) in tertiary care settings can be challenging, as patients often present with complex symptom profiles and false positives are common. We used machine learning (ML) to test our composite screening application--EarlyDetect (ED), which includes extant validated scales as well as a selection of life history factors--for predictive accuracy, with the central aim of improving specificity in tertiary settings.
Methods
This was a retrospective, naturalistic study at a tertiary mental health centre in western Canada. Participants (n = 955; 56.4% female; mean age 35.4; 34.6% MDD) completed ED and underwent a clinical interview with a blinded psychiatrist for diagnostic accuracy. ML was used to make more confident predictions at an individual level.
Results
Using composite scoring, the balanced accuracy of our tool was 72.0%, with a sensitivity of 74.2% and a specificity of 69.8%, with an AUC of 0.781. Compared with the MDD symptoms-only model, the fully composite ED model improved balanced accuracy by 5.6%, specificity by 13%, and AUC by 0.072.
Limitations
Patients were assessed using clinical psychiatric evaluations, which are subjective. There is also the potential for self-reporting bias, particularly with depressed patients. Finally, the cross-sectional design of this study rules out conclusions of causality.
Conclusions
Our results show improved MDD. detection accuracy using composite measures and highlight key predictive factors that contribute to the accurate diagnosis of MDD, including disability, family history of mental illness, and stressful events. These factors may need to be included in regular screening assessments.
Language: en |