Predictive Modelling of Depression Treatment Response using Individual Symptoms and Latent Factors

Machine learning models have increasingly been used to identify predictors of treatment response in depression, and it is hoped that they may eventually help with clinical decision making. However, the performance of these models has generally been poor. One possible reason is that they are typically trained to predict aggregate scores of several depression symptoms; by contrast, individual symptoms may behave differently, be more predictable and/or more responsive to treatment. We tested this possibility by comparing the performance of machine learning models for predicting early response to psychotherapy based on 21 different outcome measures: (i) 16 individual depression symptoms, (ii) 4 latent symptom factors for sleep, appetite, motivation, and negative affect related symptoms, and (iii) total scores based on the widely used Quick Inventory of Depressive Symptomatology (QIDS). We used a large real-world dataset of 85 baseline features spanning sociodemographic, cognitive, clinical, lifestyle and physical health assessments in patients (N=776) initiating internet-delivered cognitive behavioural therapy (iCBT). For all 21 outcome measures, we developed elastic net models (N=543) and validated their performance in an unseen hold-out sample (N=233). In the hold-out dataset the model predicting total depression scores achieved an R ² of 40% variance explained, while there was substantial variability in model performance for individual symptoms (R ² :2.1%-44%) and latent symptom factors (R ² :26%-44%). Model comparisons revealed that most individual symptom and latent factor models with all 85 predictors were not superior to simpler benchmark models comprising only age, sex and baseline levels of the respective depression outcome measure. The benchmark was outperformed by models predicting total scores ( ΔR ² =0 . 054, p=0 . 034) , sad mood ( ΔR ² =0 . 106, p=0 . 001) , loss of interest ( ΔR ² =0 . 079, p=0 . 021) and a latent factor representing negative affect and thought ( ΔR ² =0 . 054, p=0 . 038) . Specifically, these models benefitted from additional predictors, such as treatment expectation, suicidal ideation, social support, or functional impairment. Our predictive modelling approach suggests new avenues towards a more patient-centred precision psychiatry, by providing clinicians with individual-level prognoses and predictors for interventions at the symptom level.