A patient sits in the clinic, hoping for guidance on how best to manage their type 2 diabetes (T2D). For their doctor, finding this ‘best’ answer involves weighing the patient’s condition and complex clinical guidelines, then deciding which treatment option will be most effective yet acceptable for the patient—all under the constraints of a brief consultation.
These complexities often lead to inadequate or delayed adjustments to treatment plans, which in turn limit disease control and clinical outcomes.
To address this challenge, Pavitra Krishnaswamy, Principal Scientist at the A*STAR Institute for Infocomm Research (A*STAR I²R), formed a multidisciplinary research team in partnership with clinicians from Singapore General Hospital’s Department of Endocrinology. Their goal was to develop an artificial intelligence (AI) tool that could help clinicians personalise treatment decisions for patients with T2D more effectively.
“With the growth in electronic health records data that link patient characteristics and prescription choices with real-world outcomes, we have new opportunities for using AI to inform more tailored treatment decisions,” said Krishnaswamy.
The team’s efforts resulted in the creation of the AI Drug mix and dose Advisor (AIDA), an innovative tool that uses electronic health records (EHR) data to generate optimal drug mix and dose recommendations based on each patient’s needs.
An architectural overview of AIDA. Given a patient’s visit medical profile, AIDA identifies potential treatment regimens, filters and ranks them by feasibility using evidence-based diabetes management guidelines, then uses a heuristic search to identify the regimen that would optimise blood sugar (HbA1c) control relative to an individualised target (T).
© A*STAR Research
AIDA was trained on EHR data from over 700,000 patient visits in the SingHealth Diabetes Registry. Given patient factors such as age, weight, test results, medical history and past prescriptions, AIDA employs a predict-then-optimise strategy to pinpoint the smallest treatment adjustments—such as increasing dose or adding a specific drug—needed to bring a patient’s blood sugar under control.
What sets AIDA apart is its ability to identify granular drug and dose adjustments that would optimise outcomes in line with clinical guidelines. “Our algorithmic approach was motivated by the need for clinical actionability and acceptability,” said Krishnaswamy.
Through evaluations on large datasets, the team estimated that AIDA recommendations would improve control of glycated haemoglobin (HbA1c) —a measure of blood sugar —over standard-of-care advice. In qualitative reviews, a panel of endocrinologists mostly rated AIDA recommendations as clinically reasonable and precise.
By providing actionable insights, the tool could serve as a co-pilot for doctors in making personalised, timely, evidence-based and effective prescriptions with greater confidence, regardless of their specialisation level.
AIDA was developed under the ‘Diabetes Clinic of the Future’ initiative, as part of the strategic partnership between A*STAR and SingHealth. The team has been building a strong intellectual property portfolio and expanding clinical evaluations to further drive this technology to translation and impact.
The A*STAR-affiliated researchers contributing to this research are from the A*STAR Institute for Infocomm Research (A*STAR I²R).
