AI at the Bedside of Psychiatry: Comparative Meta-analysis of Imaging vs Non-imaging Machine-Learning Models for Differentiating Bipolar vs Unipolar Depression at First Episode v2

First-episode bipolar disorder (BD) is frequently misclassified as unipolar major depressive disorder (MDD). Artificial intelligence/machine-learning (AI/ML) models trained on imaging (e.g., MRI, EEG) and non-imaging (e.g., EHR/clinical, voice/NLP, actigraphy, blood/proteomic biomarkers) data show promise for early differentiation. This protocol describes a PRISMA-aligned systematic review and meta-analysis comparing diagnostic performance across modalities. We include primary human studies (2014–2025) that report test-set discrimination (primary outcome: AUC; secondary: sensitivity/specificity at a specified threshold and calibration metrics). Risk of bias will be appraised with QUADAS-2 (diagnostic accuracy framing) and PROBAST/AI considerations (prediction-model analysis). Random-effects synthesis will use Hartung–Knapp adjustment with REML for τ². A priori, we contrast imaging vs non-imaging models and run sensitivity analyses excluding high-risk-of-bias or potential data-leakage studies. Search snapshot: identified 158 records, removed 39 duplicates, screened 119, included 17 (qualitative), of which 6 were meta-ready.