Harnessing AI to Determine Safe Discontinuation of Antidepressants

Researchers at the University of South Australia have developed innovative machine learning models that hold promise for guiding clinicians on when it is safe for patients to stop long-term antidepressant medication. By analyzing dispensing data from the Pharmaceutical Benefits Scheme (PBS), the team identified patterns associated with successful medication withdrawal among over 100,000 patients prescribed antidepressants over a decade.

The study, titled "Predicting Antidepressant Deprescription with Machine Learning Using Administrative Data," was showcased at MedInfo 2025, a prominent conference focused on digital health and informatics. The findings are also published in the journal "Studies in Health Technology and Informatics."

With antidepressant use increasing worldwide—especially in countries like Australia, Iceland, Portugal, Canada, and the UK—this AI breakthrough may assist general practitioners in safely reducing or discontinuing medication that is no longer clinically necessary. Long-term use of these drugs can lead to side effects such as weight gain, sexual dysfunction, and cardiac issues, emphasizing the importance of appropriate deprescription.

However, stopping antidepressants can be challenging, as about half of patients experience withdrawal effects. Dr. Lasantha Ranwala, a medical practitioner and AI researcher, explains that clinicians are often hesitant to cease prescriptions due to concerns about withdrawal. The AI models address this by identifying patient patterns that indicate a higher likelihood of successful discontinuation.

Successful deprescription was defined as a patient remaining off antidepressants for at least a year post-discontinuation, with unsuccessful cases marked by an increase in dosage within six months. The study trained two machine learning algorithms: one based on the final prescription records, which achieved 81% accuracy, and another tracking the entire course from initial prescription to dose reduction, which achieved 90% accuracy.

These results suggest that AI-driven analysis of administrative health data can serve as a valuable decision-support tool, providing clinicians with more confidence when planning deprescription. The team aims to refine the models further and explore their practical application in clinical settings, ultimately improving medication management for patients.

This advancement underscores the potential of passive health data collection and AI to enhance personalized medicine and optimize treatment strategies in mental health care.