438-P: Prediction of Heart Failures for Diabetics Only Using Major Longitudinal Lab Test Results

Background: In the presence of a large data set of electronic health records (EHRs), predicting the future diabetic complications is of importance for decision making in the medical treatments. Using modern machine learning techniques, it is generally becoming easier to build complex models to predict the future. While complex models are giving good prediction performance, simple models should be more useful in terms of interpretability and practicality in the real medical fields. For example the less the explanatory variables such as lab tests are used for the model, the more useful it is. Our interests thus lie in whether accurate prediction is possible when using less and major explanatory variables to make a model to predict the future diabetic complications.

Method: In this paper, we make a prediction model of heart failures for diabetics in half, 1, 3, 5 years using 13 longitudinal lab tests records including hemoglobin A1c, fasting blood glucose, low-density cholesterol, high-density cholesterol, triglycerides, uric acid, serum creatinine, estimated glomerular filtration rate, proteinuria, albuminuria, gamma-glutamyl transpeptidase, alanine transaminase, aspartate transaminase, from which we extracted several longitudinal statistics for input variables. We also compared the prediction performance using several major machine learning algorithms.

Results: In prediction of future heart failures in half, 1, 3, 5 years, AUC of 0.77, 0.79, 0.79, and 0.80 were marked when using Logistic Regression. Comparing algorithms for half year prediction, ensemble algorithms outperformed Logistic Regression. The AdaBoost algorithm marked AUC of 0.82.

Conclusion: We observed that when using only 13 longitudinal lab tests records, future heart failure prediction is successfully made by using modern machine learning algorithms.