Predicting diabetes related complications in Alberta,Canada for health jurisdictions:A machine learning prediction approachOA

Background:Diabetes-related complications pose significant burdens on healthcare systems,necessitating effective predictive tools for early intervention.In the Canadian healthcare context,leveraging population-level data to develop machine learning models for predicting long-term diabetes complications remains a critical need.Methods:A prognostic study was conducted using population-level administrative health data from Alberta,Canada(2013-2021).The cohort included Albertans newly managed with diabetes medications.A tree-based machine-learning model was developed using data from 2013-2017 and temporally validated on data from 2018-2021.Performance was evaluated using discrimination,calibration,and clinical utility metrics.A cost-savings analysis was also conducted using the full 2013-2021 dataset.Results:The final cohort comprised 382,129 participants.Pre-test probabilities were low,ranging from 0.0045(retinopathy)to 0.13(cardiovascular).Due to this high-class imbalance,standard accuracy was within the range from 91.55%to 99.80%in all outcomes but not an informative metric.C-statistics ranged from 0.74 to 0.85 across complications.The model provided minimal predictive information for amputation with positive LRs<10.Net benefit analysis indicated no additional clinical utility for any of the complications.True positive predictions in the top 5th percentile of predicted risk represented over$100 million in potential cost savings if complications were prevented.Interpretation:The ML model demonstrated strong discrimination for cardiovascular and dysglycemic events but limited accuracy for retinopathy,amputation,and tissue infection.Calibration was adequate for most outcomes except amputation and retinopathy.This model shows modest potential for ML-assisted intervention in identi-fying high-risk patients for cardiovascular and dysglycemic events,its utility for other complications is constrained.