Diabetol Int. 2026 Apr;17(2):
24
Naoki Sakane,
Ken Kato,
Sonyun Hata,
Erika Nishimura,
Rika Araki,
Kunichi Kouyama,
Masako Hatao,
Yuka Matoba,
Yuichi Matsushita,
Masayuki Domichi,
Akiko Suganuma,
Seiko Sakane,
Takashi Murata,
Fei Ling Wu.
Aims/introduction: Severe hypoglycemia (SH) is a major complication in adults with type 1 diabetes mellitus (T1DM). The multifactorial etiology of T1DM highlights the need for predictive tools that integrate clinical, behavioral, and technological factors. This study aimed to develop and evaluate machine learning (ML) models for predicting SH by incorporating hypoglycemia problem-solving ability, diabetes technology, and continuous glucose monitoring (CGM) indices.
Materials and methods: We analyzed data from 247 adults with T1DM (mean age 50.4 ± 13.7 years; 38.1% male; glycosylated hemoglobin 7.7 ± 0.9%) from the FGM-Japan study. A total of 22,517 feature-model combinations were evaluated across 11 ML algorithms, including logistic regression, L1-regularized regression, random forest, LightGBM, XGBoost, SVM, Naïve Bayes, SGD, neural networks, and k-nearest neighbors. Eleven candidate predictors included impaired awareness of hypoglycemia (IAH), diabetic peripheral neuropathy (DPN), CSII, rtCGM, and seven domains of hypoglycemia problem-solving ability. The model performance was assessed with fivefold cross-validation using the receiver operating characteristic-area under the curve (ROC-AUC), accuracy, precision, recall, and F1 score. Class imbalance was addressed using SMOTE.
Results: The mean ROC-AUC across models was 0.64 (range: 0.151-0.916). The average accuracy was 0.90, but the precision and recall were consistently low, with a mean recall of 0.08. The high-performing models (ROC-AUC > 0.90) were primarily Random Forest and LightGBM, which frequently incorporated domains such as problem perception, identifying problem attributes, seeking preventive strategies, evaluating strategies, and immediate management. factors. Tree-based models significantly outperformed logistic regression, Naïve Bayes, SVM, and SGD (adjusted p < 0.001), whereas the differences among the tree-based algorithms were not clinically meaningful.
Conclusions: Tree-based ML models demonstrated superior discriminative ability for predicting SH in patients with T1DM. Hypoglycemia problem-solving ability was the strongest predictor, underscoring the importance of integrating behavioral self-management skills with clinical and technological factors.
Trial registration: University hospital Medical Information Network (UMIN) Center: UMIN000039475), Approval date 13 February 2020.
Keywords: Hypoglycemia problem-solving; Impaired awareness of hypoglycemia; Severe hypoglycemia; Type 1 diabetes