An artificial-intelligence medication app correctly classified 94% of the activity clips set aside for testing, but the clinical trial behind it involved only eight people over 15 days. The early results suggest the system can be operated in a hospital pilot, not that it improves adherence, quality of life or other health outcomes.
The system was designed to support what the researchers call rational drug use. It sends medication reminders, asks patients to record short videos while taking medicine, and sends those videos to a server that classifies the recorded action. A physician web platform and a patient mobile app make up the client side of the system.
What the system learned to recognize
For model development, 31 volunteers supplied 357 original videos, which were turned into 1,626 short clips. The clips covered three target actions, taking medicine, drinking water and eating, alongside 11 unrelated activities used as negative examples. The recordings were divided into training, validation and test sets in a 60:20:20 split, with the test set reserved for the final assessment.
The model, called PHAR in the study, combined object detection with human-activity recognition. It used convolutional and temporal computer-vision methods to examine 16 ordered video frames, with three additional focus views. On the test set, accuracy was 94%, while reported precision, recall and F1 scores were all above 0.90. The three-dimensional motion-feature version performed better than the two-dimensional version particularly when identifying medicine-taking.
Those figures describe recognition of recorded actions, not proof that people took their medicines correctly in everyday life. The study did not report exact class-by-class results, confidence intervals or a confusion matrix, so the headline accuracy gives only a broad picture of how the model performed on the researchers' test data.
A workable short-term test
The hospital evaluation was a prospective observational feasibility pilot, meaning it was set up to examine whether the system could be used and accepted rather than to test clinical effectiveness. Eight adults took part for 15 days. All completed the study, successful video uploads exceeded 90%, and the researchers described application adherence, medication verification and usability as favorable.
The wider study had several stages. A preliminary survey included 300 patients, while the image library was built with 31 volunteers. Eight people were then selected for the hospital pilot. Participants had to be at least 18, able to take continuous medication on their own and comfortable with basic technology; people with cognitive, sensory or functional limitations, among other exclusions, were not included.
The pilot group also had signs of a practical need for medication support. Half said they often forgot to take medicines, a quarter received help with drug use, and 87.5% said they discarded one to three boxes while keeping unfinished boxes. Three-quarters expected the application to be beneficial. The study's image-library participants had a mean age of 44 years, and only 6.5% met the study's definition of polypharmacy, taking four or more medicines.
Participants liked the idea, but the evidence remains early
At the end of the pilot, 75% of participants said the application supported rational drug use. A reported 87.5% said the reminders were effective and that the system helped them manage medicines independently. The same proportion said links with public and private hospitals, pharmacies and e-Nabiz would be useful.
The researchers also compared health-related quality-of-life scores at the start and after 15 days. General health rose from 70.6 to 71.8, but that change was not statistically significant, with p=0.104. Scores for vitality or energy, mental health or emotional well-being, and social functioning showed statistically significant exploratory improvements, with reported p-values of 0.001, 0.031 and 0.019. Other domains did not show significant changes.
The quality-of-life findings need particular caution. There was no concurrent control group, and the pilot was too small and short to separate a system effect from ordinary fluctuation, regression to the mean, placebo effects or the extra attention participants received during the study. The authors describe these results as preliminary and hypothesis-generating, and say larger controlled studies are needed to establish clinical effectiveness.
What a larger study would need to answer
The study does not show that the app improves medication adherence or clinical outcomes, or works better than usual care. It also does not establish long-term engagement, false-alert rates, interoperability, authentication or data-protection performance in routine practice. Because people with cognitive impairment were excluded and polypharmacy was uncommon, the results may not represent older adults or patients taking many medicines.
The next test would need more participants, a control group, longer follow-up and objective measures of whether medicines were actually taken. It would also need to include older adults and people with polypharmacy or cognitive impairment. For now, the findings offer preliminary technical and usability evidence for digital medication monitoring, while leaving its clinical value unanswered.
The work was funded by the Scientific and Technological Research Council of Türkiye under project 7218037. The authors declared no conflicts of interest, and the study data are available from the corresponding author on reasonable request.
Paper data and sources
Original title: Rational Drug Use Support System: design, artificial intelligence integration, and pilot feasibility trial
Authors: Kahraman M, Bozkurt M, Kotan S et al.
Journal/Repository: Revista da Associacao Medica Brasileira (1992)
Status: Peer-reviewed
First online: 2026-10-19
DOI: Not available
Original paper · Full text