For clinicians and patients, we hope that such symptom-level predictive models allow more personalised prognoses and insights on how to improve treatment outcomes (e.g., focusing on expectations and social support) for the symptoms that are most debilitating for a given patient!

For future precision psychiatry research, we hope this work shows the promise of incorporating symptom-derived latent factor analysis into predictive modelling & inspires new computational avenues for capturing phenotypic variation in depression to improve treatment prediction!

In terms of model explainability, our choice of model (elastic net) allowed us to examine significant predictor contributions. Interestingly, besides baseline scores of the predicted measure, treatment expectations and social support were amongst the strongest predictors!

Most interesting result: Lots of performance variability across symptoms, but our latent factor model, capturing core symptoms like negative affect & thought, was the overall winner during external validation (light blue bars) and even outperformed the total depression scores!

As predictors, we had a range of multimodal measures, including sociodemographic, cognitive, clinical, lifestyle and physical health data from real-world treatment seeking patients.

We developed and compared models predicting early response (4 weeks) to psychotherapy based on: (i) 16 individual depression symptoms, (ii) 4 latent symptom factors for sleep, appetite, motivation and negative affect related symptoms, and (iii) total scores.

Machine learning (ML) models are increasingly popular clinical support tools but typically trained to predict an aggregate score of several depression symptoms; by contrast, individual symptoms may behave differently, be more predictable and/or more responsive to treatment.

Read on for more details and cool results...

Short Version: We developed ML models to predict depression total scores, individual symptoms, and latent symptom-derived factors after a 4-week psychotherapy intervention and validated in unseen hold-out samples for generalisability and treatment-specificity (to antidepressants).