Background <p>Bipolar Disorder (BD) is a serious mental illness characterized by recurrent episodes of mania and depression. Early detection and the prediction of shifts between mood states are critical for effective management and suicide prevention. Machine learning models analyzing data from clinical records, neuroimaging, and passive sensing (e.g., smartphones) are emerging as promising tools for this purpose.</p> Objective <p>This systematic review evaluates the use of machine learning for the early detection of BD and the prediction of mood shifts, synthesizing evidence on different data modalities, algorithms, and predictive performance.</p> Methods <p>A systematic search of PubMed, PsycINFO, and IEEE Xplore was conducted for studies published up to January 2026. We included original research articles that applied machine learning to data from individuals with BD for diagnosis/classification or prediction of mood episodes.</p> Results <p>A total of 37 studies met the inclusion criteria. For early detection (BD vs. HC or MDD), neuroimaging-based models achieved mean AUCs of 0.82–0.88. For mood shift prediction, digital phenotyping models using smartphone sensor data achieved AUCs of 0.75–0.85 for predicting manic or depressive episodes 1–2 weeks in advance.</p> Conclusion <p>Machine learning shows significant promise for both early detection and mood shift prediction in BD. Multimodal approaches combining neuroimaging, clinical, and digital phenotyping data represent the most promising future direction. However, larger, prospective studies with external validation are needed before clinical implementation.</p>