Artificial Intelligence for Population -Level Prediction and Prevention of Type 2 Diabetes

Abstract

Background: Extensive implementation of electronic health records (EHRs) in the UK is a good chance to implement machine learning to predict diseases at a population level. T2DM has become one of the key health concerns of the population. Though the NHS Diabetes Prevention Programme (NHS DPP) proved to be effective in real-life situations, it is pivotal that the interventions should be directed at individuals who are at the greatest risk in order to be cost-effective. The present research uses UK healthcare as the source of data to train and test a machine learning model on predicting the occurrence of T2DM and to determine the relative significance of the risk factors at various preclinical stages Methods: Our study was a retrospective cohort study on UK Biobank data and primary care EHR (e.g., CPRD) data in 2020-2024. Our group consisted of adults who did not have underlying diabetes. The main finding was a confirmed diagnosis of T2DM, which is a combination of diagnostic codes (Read Codes / ICD-10) and diabetes medication prescriptions (excuding Metformin used in the treatment of prediabetes) and the level of HbA1c (.≥6.5%). Our L1-regularised logistic regression model was developed with a large feature set, which included demographics, clinical diagnoses, procedures, prescribed medications, and time-varying trend in laboratory values. The positive predictive value (PPV) and the area under the curve of receiver operating characteristics (AUC) were used to assess model performance. Results: The superior machine learning model was much better than the parsimonious model using traditional risk factors. The enhanced model created an AUC of 0.82 and the baseline model created an AUC of 0.76 to predict T2DM two years into the future. PPV of the top 1,000 persons who were determined to be high-risk was 24% of the enhanced model and 12% of the parsimonious model. Major predictors were high HbA1c, triglycerides and blood glucose. New risk factors or those that were recently discovered were sleep apnea, chronic liver disease, and the use of certain medications (e.g., statins, corticosteroids). These factors had a higher predictive value among the younger adults (below 50), than the older population. Conclusion: The application of machine learning to UK EHR data will offer a scalable, powerful tool to population-level risk stratification of T2DM. This would be a major red flag on the conventional approaches, and more targeted preventative measures such as the NHS DPP could be approached. The specified risk factors, depending on the age and being close to the diagnosis, provide useful insights to the creation of the clinical hypothesis and individualized prevention measures.

Ähnliche Arbeiten