Identification of severe hypoglycemia in adults with type 1 diabetes using CGM-based machine learning: evidence from the FGM-Japan study.
Severe hypoglycemia (SH) in adults with type 1 diabetes mellitus (T1DM) is associated with significant morbidity and mortality; however, its underlying causes are often complex and multifactorial. Improved tools to identify individuals at a high risk of SH are critically needed. In this study, machine learning techniques were applied to continuous glucose monitoring (CGM) data to identify distinguishing features between individuals with and without SH episodes.
We analyzed data from the real-world study of adults with T1DM enrolled in the FGM-Japan study. Eleven machine learning algorithms using continuous glucose monitoring (CGM) metrics were applied to identify SH and assess the relative importance of the contributing features. The CGM metrics included mean glucose/GMI, time above range (TAR > 250 and > 180 mg/dL), time in range (TIR 70-180 mg/dL), time below range (TBR < 70 and < 54 mg/dL), coefficient of variation (%CV), and glycemic risk index (GRI). The model performance was evaluated using the area under the receiver operating characteristic curve (AUC), accuracy, precision, recall, and F1 score.
Data from 264 adults with T1DM were analyzed. Across the models, XGBoost showed the highest AUC, significantly outperforming logistic regression, k-NN, and SVM but performed marginally below Naive Bayes. The F1-score analysis showed that logistic regression and neural networks provided a better balance between precision and recall. The model using four CGM variables (TBR < 70, %CV, GMI, and GRI) achieved the highest AUC of 0.794.
XGBoost offers strong overall discrimination; however, simpler models exhibit better F1 performance. Features like 'TBR', '%CV', 'GMI,' and 'GRI' were key features, suggesting their usefulness in identifying individuals at risk for adverse glycemic events.
Clinical Trial Registry No. UMIN000039376.
The online version contains supplementary material available at 10.1007/s13340-025-00872-4.
We analyzed data from the real-world study of adults with T1DM enrolled in the FGM-Japan study. Eleven machine learning algorithms using continuous glucose monitoring (CGM) metrics were applied to identify SH and assess the relative importance of the contributing features. The CGM metrics included mean glucose/GMI, time above range (TAR > 250 and > 180 mg/dL), time in range (TIR 70-180 mg/dL), time below range (TBR < 70 and < 54 mg/dL), coefficient of variation (%CV), and glycemic risk index (GRI). The model performance was evaluated using the area under the receiver operating characteristic curve (AUC), accuracy, precision, recall, and F1 score.
Data from 264 adults with T1DM were analyzed. Across the models, XGBoost showed the highest AUC, significantly outperforming logistic regression, k-NN, and SVM but performed marginally below Naive Bayes. The F1-score analysis showed that logistic regression and neural networks provided a better balance between precision and recall. The model using four CGM variables (TBR < 70, %CV, GMI, and GRI) achieved the highest AUC of 0.794.
XGBoost offers strong overall discrimination; however, simpler models exhibit better F1 performance. Features like 'TBR', '%CV', 'GMI,' and 'GRI' were key features, suggesting their usefulness in identifying individuals at risk for adverse glycemic events.
Clinical Trial Registry No. UMIN000039376.
The online version contains supplementary material available at 10.1007/s13340-025-00872-4.
Authors
Sakane Sakane, Hirota Hirota, Yamamoto Yamamoto, Miura Miura, Takaike Takaike, Hoshina Hoshina, Toyoda Toyoda, Saito Saito, Hosoda Hosoda, Matsubara Matsubara, Tone Tone, Kawashima Kawashima, Sawaki Sawaki, Matsuda Matsuda, Domichi Domichi, Suganuma Suganuma, Sakane Sakane, Murata Murata
View on Pubmed