Validation of Machine Learning-Based Individualized Treatment for Depressive Disorder Using Target Trial Emulation

Area of Science:

• Psychiatry
• Medical Informatics
• Pharmacology

Background:

• Depressive disorder treatment often involves trial-and-error, leading to suboptimal outcomes.
• Personalized medicine approaches are needed to improve treatment efficacy for depression.
• Existing prediction models for treatment response in depression have limitations.

Purpose of the Study:

• To develop and validate machine learning (ML) prediction models for selecting individualized pharmacological treatments for patients with depressive disorder.
• To assess the effectiveness of an individualized treatment strategy based on ML predictions in reducing treatment failure.
• To emulate clinical trials for evaluating the real-world impact of ML-guided treatment selection.

Main Methods:

• Utilized data from Taiwan's National Health Insurance Research Database, including patients with incident depressive disorders.
• Defined treatment failure as psychiatric hospitalization, self-harm hospitalization, emergency visits, or treatment change.
• Trained Super Learner ensemble prediction models for initial and next-step treatments, developing an individualized strategy to select drugs with the lowest predicted failure probability.

Main Results:

• The Super Learner model achieved an area under the curve (AUC) of 0.627 for initial treatment and 0.751 for next-step treatment.
• Model-selected regimens were associated with significantly reduced treatment failure rates: a 0.84-fold decrease for initial treatment and a 0.82-fold decrease for next-step treatment.
• Emulated clinical trials confirmed that the ML-guided individualized treatment strategy reduced treatment failure rates.

Conclusions:

• Machine learning-based prediction models can effectively guide individualized pharmacological treatment selection for depressive disorder.
• An individualized treatment strategy using ML predictions is associated with a lower probability of treatment failure.
• This approach holds promise for optimizing depression management and improving patient outcomes.