Machine learning for the diagnosis accuracy of bipolar disorder: a systematic review and meta-analysis

Background: Diagnosing bipolar disorder poses a challenge in clinical practice and demands a substantial time investment. With the growing utilization of artificial intelligence in mental health, researchers are endeavoring to create AI-based diagnostic models. In this context, some researchers have sought to develop machine learning models for bipolar disorder diagnosis. Nevertheless, the accuracy of these diagnoses remains a subject of controversy. Consequently, we conducted this systematic review to comprehensively assess the diagnostic value of machine learning in the context of bipolar disorder. Methods: We searched PubMed, Embase, Cochrane, and Web of Science, with the search ending on April 1, 2023. QUADAS-2 was applied to assess the quality of the literature included. In addition, we employed a bivariate mixed-effects model for the meta-analysis. Results: 18 studies were included, covering 3152 participants, including 1858 cases of bipolar disorder. 28 machine learning models were encompassed. Sensitivity and specificity in discriminating between bipolar disorder and normal individuals were 0.88 (9.5% CI: 0.74 similar to 0.95) and 0.89 (95% CI: 0.73 similar to 0.96) respectively, and the SROC curve was 0.94(95% CI: 0.92 similar to 0.96). The sensitivity and specificity for distinguishing between bipolar disorder and depression were 0.84 (95%CI: 0.80 similar to 0.87) and 0.82 (95%CI: 0.75 similar to 0.88) respectively. The SROC curve was 0.89 (95%CI: 0.86 similar to 0.91). Conclusions: Machine learning methods can be employed for discriminating and diagnosing bipolar disorder. However, in current research, they are predominantly utilized for binary classification tasks, limiting their progress in clinical practice. Therefore, in future studies, we anticipate the development of more multi-class classification tasks to enhance the clinical applicability of these methods.