Introduction People with type 2 diabetes mellitus (T2DM) are at increased risk of developing heart failure (HF). Machine learning (ML) has the advantage of computing several patient characteristics on a time-to-event basis, compared to conventional statistics. In a general population setting, we identified factors that predicted incident HF in patients with T2DM.

Methods National Health Service Scotland electronic medical records (EMRs) were linked with the Scottish Care Information - diabetes registry including demographic data, routine laboratory measurements, prescriptions, death records and comorbidities from primary and secondary care diagnostic codes. Incident HF was defined by the International Classification of Diseases, 10th Revision (ICD-10) codes for hospitalisations, with a look-back period of 5 years to exclude prevalent cases. We developed an extension of the random forest model: a nonparametric decision tree, which supports time-to-event data, to predict incident HF. We used Cox proportional hazards models to investigate associations between prescription of loop diuretic and risk of new onset heart failure. We applied a state-of-the-art ML explainability method called Shapely Additive Explanations which interprets the direction of association for each factor on the model's prediction.

Results Out of 30,495 patients with T2DM age >50 years, 1,476 (5%) had incident HF between 2009-2019. Table 1 shows the key factors predicting incident HF: use of loop diuretics, history of atherosclerosis events (myocardial infarction, angina and atrial fibrillation), lower estimated glomerular filtration rate (eGFR ) and older age. Individuals prescribed loop diuretics, had a 5-fold higher risk of incident HF than those who were not (HR: adjusted for age and sex 5.89 [95% CI 5.27 – 6.58 (<0.005)]). People with greater socioeconomic deprivation were also at greater risk. The model c-statistic score was 0.82 and the brier score was 0.03 (low values indicate greater accuracy) for predicting incident HF.

Conclusion Incident HF risk trajectories are known to vary widely in a general population of patients with T2DM. Use of loop diuretics is an important marker to identify those at greater risk.

Conflict of Interest n/a