Predicting preventative effects of cognitive control training in remitted depressed individuals: A machine learning approach

Abstract

Background

Residual cognitive complaints are frequently observed in remitted depressed individuals (RMD), which can impair full recovery and increase the likelihood of recurrent episodes of depression. Cognitive control training (CCT) has shown potential as a preventative intervention in RMD with small to moderate effect sizes, but substantial heterogeneity in effects between individuals exists.

Objective

This study aimed to identify individual characteristics associated with CCT treatment response in RMD participants using machine learning (ML) models.

Methods

227 RMD underwent a CCT regimen of at least 10 sessions. Three machine-learning models were evaluated: logistic regression, random forest, and XGBoost, alongside one random classifier benchmark. Performance metrics (accuracy, precision, recall, F1-score, and AUC-ROC) were computed. Feature importance was assessed using SHAP values.

Result

All models demonstrated low performance, regardless of ML methodology. The logistic regression model obtained the highest performance, although this was still considered low (accuracy: 0.54; AUC-ROC: 0.49). Exploratory feature importance analysis revealed that age, sense of well-being, and life satisfaction were important variables in the models, while current use of psychotherapy, number of prior depressive episodes, and history of inpatient treatment were not.

Conclusion

All models performed poorly, indicating that baseline characteristics did not confidently predict CCT treatment effects in this RMD sample. Exploratory feature analysis indicates that some clinical variables may increase the likelihood of benefiting from CCT, while most demographical variables did not seem to affect treatment effectiveness. However, due to low model performance, confidence in feature importance was low and additional research using larger samples is required.