AI models predict when patients can safely stop long-term antidepressant use

New machine learning models developed by University of South Australia (UniSA) researchers could help clinicians identify when patients can successfully stop long-term antidepressant use.

Using artificial intelligence (AI) to track dispensing data from the Pharmaceutical Benefits Scheme (PBS), UniSA researchers identified the most successful cases of deprescription among 100,000 patients prescribed antidepressants over a 10-year period.

The study, “Predicting Antidepressant Deprescription with Machine Learning Using Administrative Data,” was presented at MedInfo 2025, an international conference on digital health and informatics. The findings are also published in Studies in Health Technology and Informatics.

With antidepressant use soaring worldwide—Australia, Iceland, Portugal, Canada and the UK top the consumption levels globally—the AI breakthrough could help general practitioners safely deprescribe medication that is no longer clinically recommended.

While these drugs can be life-changing, prolonged use is linked to a range of side effects, including weight gain, sexual dysfunction and cardiac issues.

However, 50% of patients also experience withdrawal effects when they cease taking them, so it’s a delicate balancing act to manage therapeutic benefits and risks, according to Dr. Lasantha Ranwala, a medical practitioner, AI researcher and UniSA Ph.D. candidate.

“Health care providers are often reluctant to cease antidepressant prescriptions due to their concerns about withdrawal effects, making it difficult for doctors to know who can safely discontinue treatment,” Dr. Ranwala says.

“By applying AI to the PBS database, we have identified patterns linked to successful withdrawal, forecasting which patients are most likely to succeed when taking them off antidepressants.”

Successful deprescription was defined as the absence of any antidepressant medication for at least one year following previous long-term use (more than 12 months). If medicine strength increased within six months following a reduction attempt, it was labeled a failure.

Researchers say the findings could give clinicians a powerful decision-support tool, helping them initiate deprescription with greater confidence.

Two machine learning algorithms were trained and tested. One assessed final prescription records (achieving 81% accuracy) and the other tracked patients from their first prescription, monitoring dose reductions and outcomes (achieving 90% accuracy).

“These results show real promise,” says UniSA co-author Associate Professor Andre Andrade.

“The most accurate model was the one that offered a more nuanced picture of deprescription attempts, better reflecting patient experiences,” Assoc Prof Andrade says.

The findings suggest that administrative health care data could help predict clinical outcomes and improve medical decision making.

“This data is passively collected, underused by medical professionals and a good candidate for AI use.”

Building on their AI tool for predicting safe antidepressant withdrawal, the researchers will now focus on making the technology more accurate and easier to use. They aim to test its effectiveness in clinics and explore how similar approaches could help patients improve their use of medicines.