4CPS-131 Development of an augmented intelligence tool to predict risk of uncontrolled type 2 diabetes mellitus for personalised pharmaceutical care at the outpatient setting

<h3>Background and Importance</h3> Over the past decades, the shrinking pharmacy workforce has gradually transformed in response to the increasing healthcare needs of the ageing population. Digital technologies such as risk stratification models using machine learning have the potential to drive towards large-scale, cost-effective and high-quality personalised pharmacy care. <h3>Aim and Objectives</h3> The aim is to develop an Augmented Intelligence (AI)-assisted risk stratification model for Type 2 diabetes mellitus (T2DM) to predict the risk of uncontrolled T2DM in the next 6 months to provide proactive and timely personalised pharmaceutical care services. <h3>Material and Methods</h3> A retrospective cohort study was conducted in T2DM patients followed-up at Singapore General Hospital Specialist Outpatient Clinic for at least 1 year from 2021 to 2024. Uncontrolled T2DM is defined as glycated haemoglobin (HbA1c) &gt;7% for patients &lt;75 years old and HbA1c &gt;8% for patients ≥75 years old. A set of features including demographic, physical measurements, laboratory data, medical history, diabetes medications, medication adherence, clinic attendances and clinical interventions were included. Random forest, logistic regression and gradient boosting classifier algorithms were employed. The area under the curve (AUC), recall, F1-measure, precision, and accuracy metrics were evaluated. <h3>Results</h3> A total of 1140 T2DM patients records were included. Table 1 summarises the performance metrics achieved by each algorithm. Both the Random Forest and Gradient Boosting Classifier models performed better than Logistic Regression across the various evaluation criteria using the top 22 features plus the medication adherence components. Baseline HbA1c, age, total number of diabetes medications, serum creatinine and triglycerides were important predictors of uncontrolled T2DM. <h3>Conclusion and Relevance</h3> Both the Random Forest and Gradient Boosting Classifier models demonstrated high performance in predicting uncontrolled T2DM. Machine learning techniques are promising to build accurate models to forecast disease outcomes and provide large-scale personalised pharmacy care. <h3>References and/or Acknowledgements</h3> <h3>Conflict of Interest</h3> Corporate sponsored research or other substantive relationships: The project was supported by the Singapore General Hospital Innovation Grant (SIG-07–2023).