# Multimodal Prediction of Future Depressive Symptoms in Adolescents

## TL;DR

This study uses a combination of clinical, behavioral, and neural data to predict future depressive symptoms in adolescents.

## Contribution

The study introduces a multimodal approach integrating multiple data types to improve short-term prediction of adolescent depression.

## Key findings

- Multivariable linear regression using principal component scores best predicted CES-D depression scores.
- Baseline MFQ scores were the best predictor for the MFQ measure of depression.
- Items assessing melancholic features were most predictive of future depressive symptoms.

## Abstract

Depression rates surge during adolescence. Early identification of youth at increased risk for depression is crucial for timely intervention and, ideally, prevention. This study aims to improve the prediction of future depressive symptoms in adolescents by using a multimodal approach that integrates relevant clinical, demographic, behavioral, and neural characteristics.

103 adolescents (ages 12–18; 72.8% female) underwent a baseline assessment including self-report questionnaires, ecological momentary assessment, a clinical interview, and behavioral and neural measures of reward responsiveness. We used nested cross-validation to compare machine learning approaches as well as conventional linear regression in predicting depressive symptoms (Center for Epidemiological Studies Depression Scale [CES-D] and the Mood and Feelings Questionnaire [MFQ]) at a 3-month follow-up.

For the prediction of CES-D depression scores, the best performing model was a multivariable linear regression using as predictors five principal component scores from a principal component analysis of baseline variables (RMSE = 6.501, R2 = 0.688). For the MFQ, the best performing model was a univariable linear regression with baseline MFQ scores as the sole predictor (RMSE = 8.054, R2 = 0.671). A factor analysis revealed that items assessing melancholic features were most predictive of future depressive symptoms.

More complex machine learning approaches did not outperform regression in predicting future depression. The integration of relevant multimodal predictors reveals which adolescent characteristics (e.g., melancholic features and physical anxiety) have a larger contribution to predicting short-term future depression. Future studies are needed with larger sample sizes and longer follow-up periods to provide a more comprehensive test of such models.