Fig. 9: Comprehensive model comparison for predicting depression-symptom change.

a–c Prediction analyses were based on a leave-two-subject-out cross-validation procedure and their significance was tested using a permutation test with 1,000 permutations for each possible left-out pair combination. a The base-model predicts mean depression symptom increases significantly above chance (out-of-sample accuracy = 67.38%, p < 0.001, R² = 0.27 adjusted R² = 0.23). The PHQ-survey score is already a significant predictor for depression symptom severity changes (p = 0.0002). b The full model containing additional behavioral-, neural- and pupil data predicts mean depression increases significantly above chance (out-of-sample accuracy = 64.36%, p < 0.001, R² = 0.37, adjusted R² = 0.28) and increases the explained variance by 11% and the adjusted explained variance by 4.3%. c The optimal model has similar prediction improvements as the full model (out-of-sample accuracy = 67.7%, p < 0.001, R² = 0.33, adjusted R² = 0.30) but contains only two parameters: locus coeruleus upregulation response (p = 0.039) and the PHQ-depression survey (p = 0.0009). Compared to the base-model, this sparse model predicts 10% more of the variance and also 10% of the adjusted variance. d–i Receiver operating characteristic (ROC) plots and area under the curve (AUC) for different combinations of measures predicting depression: (d) Base-model, (e) Full-model, (f) Optimal model, (g) LC-only, (h) LC-Amygdala only, (i) pupil only. Please see Supplemental Table S8 for additional models, full details on single regressor contributions and model comparison. Source data are provided as a Source Data file.