Machine Learning-Based Prediction of Illness Course in Major Depression: the Relevance of Risk Factors

Background: Major depressive disorder (MDD) comes along with an increased risk of recurrence and poor course of illness. Machine learning has recently shown promise in the prediction of mental illness, yet models aiming to predict MDD course are still rare and do not quantify the predictive value of established MDD recurrence risk factors. Methods: We analyzed N = 571 MDD patients from the Marburg-M & uuml;nster Affective Disorder Cohort Study (MACS). Using random forest classifiers, we predicted i) recurrence of depressive episodes and ii) MDD disease trajectories, within a 2-year interval. Trajectories were identified through latent profile analysis, using a discovery and an internal validation sample. Three distinct models were implemented for predictions: two incorporating only literature-derived MDD recurrence risk factors, and a third incorporating a broader set of explorative features. Results: Basing predictions on only seven recurrence risk factors, MDD recurrence could be predicted with a balanced accuracy (BACC) of 62.83 % and MDD trajectories were predicted with highest performance achieved for a remitted (BACC = 64.23 %) and a severe MDD trajectory (BACC = 63.17 %). Risk factors included childhood maltreatment, previous depressive episodes, residual symptoms, comorbid anxiety, age of onset, depression severity, and neuroticism. Including a broader feature set only yielded in minor increase of predictive performance. Limitations: Lacking external validation, generalizability to other samples remains uncertain. Conclusions: MDD recurrence and disease trajectories can be predicted based on literature-derived recurrence risk factors. Model performance must increase to be of use in clinical practice which could be achieved by including multimodal risk factors.