Predicting the naturalistic course of depression from a wide range of clinical, psychological, and biological data: a machine learning approach

Abstract

Many variables have been linked to different course trajectories of depression. These findings, however, are based on group comparisons with unknown translational value. This study evaluated the prognostic value of a wide range of clinical, psychological, and biological characteristics for predicting the course of depression and aimed to identify the best set of predictors. Eight hundred four unipolar depressed patients (major depressive disorder or dysthymia) patients were assessed on a set involving 81 demographic, clinical, psychological, and biological measures and were clinically followed-up for 2 years. Subjects were grouped according to (i) the presence of a depression diagnosis at 2-year follow-up (yes n = 397, no n = 407), and (ii) three disease course trajectory groups (rapid remission, n = 356, gradual improvement n = 273, and chronic n = 175) identified by a latent class growth analysis. A penalized logistic regression, followed by tight control over type I error, was used to predict depression course and to evaluate the prognostic value of individual variables. Based on the inventory of depressive symptomatology (IDS), we could predict a rapid remission course of depression with an AUROC of 0.69 and 62% accuracy, and the presence of an MDD diagnosis at follow-up with an AUROC of 0.66 and 66% accuracy. Other clinical, psychological, or biological variables did not significantly improve the prediction. Among the large set of variables considered, only the IDS provided predictive value for course prediction on an individual level, although this analysis represents only one possible methodological approach. However, accuracy of course prediction was moderate at best and further improvement is required for these findings to be clinically useful.

Fig. 1. Model predictions.

Confusion matrices for classifiers are depicted in panel a for binary prediction, i.e., presence or absence of a unipolar depression diagnosis at follow-up (major depressive disorder or dysthymia), and b for prediction of the three LCGA course trajectory groups. Number and color in each cell describe the proportion of predictions. For example, chance level would be 0.5 in each cell in the confusion matrix in a, and 0.333 in the confusion matrix in b. Violin plots of the spread of predicted values are depicted in panel c for binary prediction, i.e., presence or absence of a unipolar depression diagnosis at follow-up, and d for predicting the three course trajectory groups

Fig. 2. Stability paths.

Stability paths of elastic-net logistic regression showing selection probabilities of each variable with respect to amount of applied regularization. The less regularization is applied, the more variables will be included in the model and the higher the chance for a false-positive selection. The stability selection approach allows us to statistically control for false-positive discovery. Variables crossing the marked regions are statistically significantly related to the outcome variable with the error correction p_fwer < 0.05 according to the stability selection theory. Other variables that crossed the probability threshold (they have been selected at least 75% of times under resampling) might also be important, but they did not survive the multiple comparison correction. a, b Logistic regression trained on all variables. c, d Logistic regression trained only on the individual items from the inventory of depressive symptomatology (IDS) questionnaire

Fig. 3. Performance of different data modalities.

Mean area under the curve for predictive models of naturalistic course of depression. a Predicting the presence or absence of a unipolar depression diagnosis 2 years after the baseline measurement. b Predicting the three depression course trajectory groups