Machine learning techniques to predict diabetic ketoacidosis and HbA1c above 7% among individuals with type 1 diabetes - A large multi-centre study in Australia and New Zealand

Abstract

Background and aims: Type 1 diabetes and diabetic ketoacidosis (DKA) have a significant impact on individuals and society across a wide spectrum. Our objective was to utilize machine learning techniques to predict DKA and HbA1c>7 %.

Methods and results: Nine different models were implemented and model performance evaluated via the Area under the Curve (AUC). These models were applied to a large multi-centre dataset of 13761 type 1 diabetes individuals prospectively recruited from Australia and New Zealand. Predictive features included a number of clinical demographic and socio-economic measures collected at previous visits. In our study, 2.9 % reported at least one episode of DKA since their last clinic visit. A number of features were significantly associated with DKA. Our results showed that Deep Learning (DL) model performed well in predicting DKA with an AUC of 0.887. The DL also provided the lowest classification error rate of 0.9 %, highest sensitivity of 99.9 % and F-measure of 99.6 %. As for HbA1c >7 %, the optimal Support Vector Machine provided a good AUC of 0.884.

Conclusion: Machine learning models can be effectively implemented on real-life large clinical datasets and they perform well in terms of identifying individuals with type 1 diabetes at risk of adverse outcomes.

Keywords: Diabetic ketoacidosis; Machine learning; Personalised medicine; Type 1 complications.