Machine learning approaches in the therapeutic outcome prediction in major depressive disorder: a systematic review

Abstract

Background: Various factors impact treatment outcomes in major depressive disorder (MDD), complicating prediction of treatment success. Therefore, applying machine learning (ML) algorithms for therapeutic outcome prediction on the basis of individual patient data has become a promising approach to tailor the treatment strategy in MDD. However, the applicability of such decision support systems in clinical settings has not been sufficiently demonstrated yet. The objective of the evaluation was to assess applicability of currently published ML-approaches for clinical settings in the EU on the basis of quality, ethical, social, and legal criteria. Methods: We performed a bibliographic search on PubMed and Google Scholar for studies from January 2016 to December 2024 on ML-applications predicting treatment outcomes in MDD. The ML-model applicability was evaluated via information on validation and performance criteria and the compliance with relevant ethical, social, and legal criteria in the EU. Results: In the 29 publications reviewed, Random Forest (RF) and Support Vector Machine (SVM) were identified as most frequently used ML-methods. Models integrating multiple categories of patient data, demonstrated higher predictive accuracy than single-category models. However, external validation of the applied ML-approaches was limited and due to the early stage of development, compliance with social, ethical and legal standards remains challenging. Conclusion: A lack of demonstrated generalizability of the evaluated ML-approaches for treatment outcome prediction in MDD and challenges with regulatory compliance in terms of relevant social, ethical and legal aspects do not yet show sufficient applicability and utility for a use in clinical settings in the EU.