Premium
Performance assessment of different machine learning approaches in predicting diabetic ketoacidosis in adults with type 1 diabetes using electronic health records data
Author(s) -
Li Lin,
Lee ChuangChung,
Zhou Fang Liz,
Molony Cliona,
Doder Zoran,
Zalmover Evgeny,
Sharma Kristen,
Juhaeri Juhaeri,
Wu Chuntao
Publication year - 2021
Publication title -
pharmacoepidemiology and drug safety
Language(s) - English
Resource type - Journals
SCImago Journal Rank - 1.023
H-Index - 96
eISSN - 1099-1557
pISSN - 1053-8569
DOI - 10.1002/pds.5199
Subject(s) - medicine , diabetic ketoacidosis , logistic regression , receiver operating characteristic , lasso (programming language) , random forest , test set , confidence interval , diabetes mellitus , area under the curve , type 1 diabetes , machine learning , type 2 diabetes , artificial intelligence , statistics , mathematics , computer science , endocrinology , world wide web
Purpose To assess the performance of different machine learning (ML) approaches in identifying risk factors for diabetic ketoacidosis (DKA) and predicting DKA. Methods This study applied flexible ML (XGBoost, distributed random forest [DRF] and feedforward network) and conventional ML approaches (logistic regression and least absolute shrinkage and selection operator [LASSO]) to 3400 DKA cases and 11 780 controls nested in adults with type 1 diabetes identified from Optum® de‐identified Electronic Health Record dataset (2007–2018). Area under the curve (AUC), accuracy, sensitivity and specificity were computed using fivefold cross validation, and their 95% confidence intervals (CI) were established using 1000 bootstrap samples. The importance of predictors was compared across these models. Results In the training set, XGBoost and feedforward network yielded higher AUC values (0.89 and 0.86, respectively) than logistic regression (0.83), LASSO (0.83) and DRF (0.81). However, the AUC values were similar (0.82) among these approaches in the test set (95% CI range, 0.80–0.84). While the accuracy values >0.8 and the specificity values >0.9 for all models, the sensitivity values were only 0.4. The differences in these metrics across these models were minimal in the test set. All approaches selected some known risk factors for DKA as the top 10 features. XGBoost and DRF included more laboratory measurements or vital signs compared with conventional ML approaches, while feedforward network included more social demographics. Conclusions In our empirical study, all ML approaches demonstrated similar performance, and identified overlapping, but different, top 10 predictors. The difference in selected top predictors needs further research.