Machine learning model non-invasively detects suspected cases of hypoglycemia during real-world driving

1. Machine learning (ML) models developed by Lehmann and colleagues used driving characteristics and head motion data to detect hypoglycemic episodes.

2. ML models detected hypoglycemia with high accuracy, even when restricted to only one of the two data inputs.

Evidence Rating Level: 2 (Good)

Study Rundown: Hypoglycemia is a persistent risk for people with type 1 diabetes, and it significantly affects cognitive and psychomotor abilities, making its occurrence during driving particularly dangerous. However, existing detection methods require invasive monitoring and are unfeasible during everyday tasks. Lehmann and colleagues conducted two studies involving individuals with type 1 diabetes. The studies collected their driving characteristics and head motion data on a secured test track. In Study 1, driving was performed under euglycemia and hypoglycemia; in Study 2, the conditions were euglycemia and mild hypoglycemia. Data from the two studies were combined and analyzed collectively. The study outcome was the diagnostic accuracy of the ML model in detecting hypoglycemia, quantified as the area under the receiver-operating characteristic curve (AUROC). The study found that the ML model achieved an AUROC of 0.80 when using both driving characteristics and head motion data. The model yielded an AUROC of 0.73 using only driving characteristics alone, and an AUROC of 0.70 using head motion alone. Overall, this study demonstrated an ML model’s high accuracy in the non-invasive detection of hypoglycemia during driving.

Click here to read the study in NEJM AI

Relevant Reading: Artificial intelligence for diabetes: Enhancing prevention, diagnosis, and effective management

In-Depth [randomized controlled trial]: People with type 1 diabetes and with glycated hemoglobin values of no more than 9.0% were included in the study. Exclusion criteria included pregnancy and comorbidities such as cardiovascular disease and seizures. Thirty participants were included in the final analysis (20 from Study 1 and 10 from Study 2). Study 1 collected data during euglycemia, defined as blood glucose (BG) between 5.0 and 8.0 mmol/L, and pronounced hypoglycemia (BG 2.0-2.5 mmol/L). Study 2 collected the same data during euglycemia and mild hypoglycemia (BG 3.0-3.5 mmol/L). Hypoglycemic states were induced via overnight fasting; driving characteristics were collected using the controller area network (CAN), and head motion data was collected using driver-monitoring cameras (DMCs). For 20 minutes in each glycemic state, the participants drove the study vehicle on a standardized test track under the supervision of a driving instructor. Three ML models were used to detect hypoglycemia: CAN + DMC, CAN only, and DMC only. The CAN + DMC ML model achieved the highest accuracy, with an AUROC of 0.80±0.11. The other two models achieved lower but respectable accuracies: 0.73±0.07 for CAN only and 0.70±0.16 for DMC only. The researchers also found that participants underestimated the effects of hypoglycemia, as 40% stated that they would have continued driving in mild hypoglycemia. Further studies are needed to assess the generalizability of this study to other groups, such as individuals with type 2 diabetes or intoxicated individuals.

Image: PD

©2024 2 Minute Medicine, Inc. All rights reserved. No works may be reproduced without expressed written consent from 2 Minute Medicine, Inc. Inquire about licensing here. No article should be construed as medical advice and is not intended as such by the authors or by 2 Minute Medicine, Inc.