Machine learning-based prediction of risk factors for abnormal glycemic control in diabetic cancer patients receiving nutrition support: a case–control study

Abstract

Purpose: To date, risk factors affecting abnormal glycemic control have not been investigated. This study aimed to analyze risk factors for hypoglycemia or hyperglycemia in diabetic cancer patients receiving nutritional support by using machine learning methods. Methods: This retrospective two-center study was performed using medical records. Odds ratios and adjusted odds ratios were estimated from univariate and multivariate analyses, respectively. Machine learning algorithms, including five-fold cross-validated multivariate logistic regression, elastic net, and random forest, were developed to predict risk factors for hypoglycemia and hyperglycemia. Results: Data from 127 patients were analyzed. The use of sulfonylurea (SU) and blood urea nitrogen (BUN) level > 20 mg/dL increased hypoglycemia by 6.3-fold (95% CI 1.30–30.47) and 5.0-fold (95% CI 1.06–23.46), respectively. In contrast, patients who received an actual energy intake/total energy expenditure (TEE) ≥ 120% and used dipeptidyl peptidase-4 (DPP-4) inhibitors had a higher risk of hyperglycemia by 19.3- (95% CI 1.46–254.78) and 3.3-fold (95% CI 1.23–8.61), respectively. An initial blood glucose level ≥ 182.5 mg/dL also increased the risk of hyperglycemia by 15.3-fold. AUROC values for all machine learning methods indicated acceptable and excellent performance for hypoglycemia and hyperglycemia. Conclusion: The use of SU and BUN level > 20 mg/dL increased the risk of hypoglycemia, whereas an initial blood glucose level ≥ 182.5 mg/dL, a supplied actual energy intake/ TEE ≥ 120%, and the use of DPP-4 inhibitors increased the risk of hyperglycemia. © 2023, The Author(s), under exclusive licence to Hellenic Endocrine Society.