Machine Learning-Based Digital Drawing Phenotyping for Exploratory Assessment of Anxiety and Depression in Patients with Inflammatory Demyelinating Diseases

Anxiety and depressive symptoms are common in patients with inflammatory demyelinating diseases (IDDs), but low-burden digital behavioral tools remain underexplored. This study examined whether digital tree drawing and person-in-the-rain tasks contain machine learning-assisted phenotyping signals associated with anxiety and depression in IDD. We analyzed 86 patients (84 for MADRS) with valid digital drawing trajectories and emotional assessments. A total of 257 demographic, trajectory, pressure, pause, content, bounding-box, and cross-task features were extracted. Continuous HAMA, HAMD, and MADRS scores and exploratory binary labels (scores of 7 or higher) were evaluated across eight prespecified candidate pipelines, defined by four feature sets crossed with ridge and elastic net models. All pipelines used identical repeated five-fold outer cross-validation splits, with preprocessing, repeated subsampling supervised univariate feature ranking, and model fitting confined to the outer training folds. Predictive performance was limited and varied by outcome. For continuous outcomes, HAMA showed modest signal across candidate pipelines (R2 = 0.082–0.133), whereas HAMD and MADRS remained close to the null baseline (R2 = −0.089 to −0.068 and −0.025 to −0.011, respectively). For binary labels, ROC-AUC ranges were 0.564–0.643 for HAMA ≥ 7, 0.590–0.609 for HAMD ≥ 7, and 0.685–0.733 for MADRS ≥ 7. Outer-fold feature-recurrence summaries most consistently identified person bounding-box dimensions and cross-task path length measures for HAMA and pause-related timing features for HAMD and MADRS. These findings suggest that digital drawing may capture limited but interpretable emotional symptom-related behavioral variation within IDD cohorts. Further multicenter and external validation is needed to evaluate its role as a complementary behavioral assessment paradigm.