Diabetes is a major global health crisis characterized by rising prevalence, severe complications, and substantial economic impact. Improper self-management, low adherence to medications, uneven medical resource distribution, and a shortage of diabetes specialists lead to poor glycemic control in diabetic patients. Recent advances in digital health technologies, specifically artificial intelligence (AI), offer a significant opportunity to attain a higher efficiency in diabetes care, diminishing the rising diabetes-related healthcare expenses.
Image Credit: Andrey_Popov/Shutterstock.com
Diabetes Management Challenges
In traditional medical practices, managing diabetes presents several challenges. For instance, early diagnosis and prevention are challenging, as several cases remain undiagnosed for several years. Similarly, the management of diabetic patients involves routine follow-up of thorough diabetic complications and blood glucose control examinations.
Additionally, collaboration among ophthalmology, nephrology, nutrition, podiatry, and endocrinology is required for integrated diabetes management. These factors lead to an uneven medical resource distribution, with an insufficient primary health care capacity and a lack of high-quality human resources.
Moreover, diabetes is one of the most highly prevalent chronic diseases that requires an active continuous role of the patient in its management owing to its extensive complications throughout the body’s physiological systems, reliance on exercise and diet, and requirement for self-monitoring.
The Need for AI
The emergence of AI could address these challenges and alleviate the diabetes disease burden in the future. Specifically, the application of machine learning has been extensively investigated in clinical practice, translational research, and basic biomedical research.
In diabetes management, AI-based digital health technologies could facilitate the implementation of effective prevention strategies for high-risk populations, deliver real-time metabolic and health information, manage diabetic patients who cannot physically attend physician appointments, save money and time by reducing travel to in-person appointments, and promote better self-management of patients.
By combining AI approaches and advanced technologies like sensor technologies, wearable devices, and medical devices, better diabetes management services could be developed and implemented. AI can be used for risk stratification, patient self-management tools, clinical diagnosis support, and automatic retinal screening in diabetes management.
In automatic retinal screening, an AI technique interprets the absence or presence of diabetic retinopathy automatically from fundus images. The Digital Diagnostics Inc.-manufactured IDx-DR device is an example of this technology. This device facilitates the diagnosis and screening of diabetic retinopathy, specifically in rural communities where patients face difficulties in accessing ophthalmologists.
In clinical diagnostic support, AI techniques that could imitate the clinical acumen of specialists, such as fine-tuning insulin doses, are being developed. FDA-approved Advisor Pro, manufactured by DreaMed Diabetes, Ltd., is an example of such an AI system. In the future, AI-based medical devices could replace diabetes specialists in tasks like fine-tuning insulin therapy.
Self-management tools are familiar to several diabetes patients as they have self-checked different biometric data, like actively measuring blood glucose levels through self-monitoring of blood glucose. The AI technology interprets the patient’s biometric data with the patient self-management tools and alerts a diabetologist to improve their blood glucose control. The Medtronic-manufactured Guardian Connect System is an example of an AI system with this functionality.
In the diagnosis and treatment of diabetes, AI technology could eliminate the prevalence of diabetes by implementing medical intervention for individuals who are highly likely to develop diabetes at a very early stage or the pre-illness stage.
AI Techniques for Diabetes Management
Various AI techniques, such as artificial neural networks (ANN), convolutional neural networks (CNN), support vector machines (SVM), random forest (RF), decision trees (DT), Bayesian networks (BN), k-means clustering, linear regression, logistic regression, machine learning classifiers, fuzzy RF, and recurrent neural networks (RNN), are utilized in diabetes management.
For instance, CNN predicts the development of chronic kidney disease, the future threat of significant diabetic retinopathy worsening at a patient level within 2 years, and the risk of developing diabetic retinopathy within 2 years using retinal images and risk factors as inputs. Similarly, DeepDR (CNN), CNN, EyeArt (CNN), morphological image processing and SVM, fuzzy RF, and IDx-DR (CNN) are effective for diabetic retinopathy screening using retinal images as input.
ML classifiers and CNN can be employed for chronic kidney disease screening using metabolic biomarkers, retinal images, and risk factors as inputs. DT, RF, SVM, and BN could screen diabetic nephropathy using clinical and genotyping data, while RF and SVM predict and differentiate diabetic nephropathy and non-diabetic renal disease using clinical data.
CNN and k-means clustering techniques classify thermograms depending on the diabetic foot complications severity using thermogram images; ANN, RF, and SVM estimate the possibility of healing of diabetes-related foot ulcers using smartphone images; and ANN, SVM, DT, linear regression, and logistic regression present AI systems for polyneuropathy screening based on electronic health records.
CNN can identify diabetic peripheral neuropathy using corneal confocal microscopy images. LASSO regression estimates the risk of amputation in patients treated with canagliflozin using risk factors as input; RNN predicts the development of 10 selected diabetic complications using hospitalization longitudinal data; and network analysis quantifies the progression of comorbidities using longitudinal data.
RF, SVM, and LR/ANN, case-based reasoning, and BN predict the development of diabetes-related complications, estimate the risk of complications, and estimate the long-term development and progression of complications, respectively, using risk factors as input.
AI Applications
Diabetes Complication Monitoring: Peripheral neuropathies and vascular pathologies are the most common diabetes complications. Studies have displayed that deep learning algorithms could effectively detect diabetic retinopathies with high specificity and sensitivity.
A novel AI disease grading system based on deep learning has been recently developed to grade the severity of diabetic retinopathies. Similarly, a fuzzy logic-based expert system has been developed to intelligently determine the diabetic neuropathy’s severity with high accuracy, specificity, and sensitivity.
In a study, researchers employed asymmetry analysis and a genetic algorithm to analyze thermal images for early foot ulceration detection and skin integrity assessment. Results showed that this proposed approach could effectively detect inflammation and predict potential foot ulcerations.
Insulin Injection Guidance: Insulin dosage is calculated using formulas with data collected from dynamic blood sugar monitoring, insulin sensitivity coefficient, and carbohydrate coefficient as inputs. AI technologies have been applied intensively to insulin injection guidance to provide enhanced insulin usage support for patients.
For instance, a closed-loop remote monitoring system based on a self-adaptive learning algorithm and fuzzy logic, MD-Logic, has been created to reduce the median nocturnal hypoglycemia duration significantly. This system is effective and can be used safely at night by type 1 diabetes patients, adolescents, and children.
Moreover, logic-based AI tools are also used for providing insulin guidance. Studies have determined the effects of exercise time, drinking time, and eating time on individual glucose metabolism using case-based reasoning. This method was employed to calculate an individualized insulin bolus using the insulin intravenous bolus calculator, thereby optimizing insulin treatment and achieving optimal glucose levels in patients.
AI-based Automatic Diet Monitoring: Dietary monitoring is crucial in diabetic patients, specifically self-reporting of food intake is often impractical and inaccurate. Thus, automated solutions for dietary monitoring are necessary.
Recent studies have demonstrated rising accuracy in energy intake estimation based on food images. For instance, a smartphone system, GoCARB, has been designed specifically for patients with type 1 diabetes to effectively estimate the carbohydrate content in meals.
Similarly, another mobile food identification system has been developed that identifies the food automatically and accurately and estimates its nutritional and caloric content without user intervention. A novel concept of “food energy distribution” was introduced in a study to model the food energy distribution characteristics in an eating scene.
A food image’s mapping to its energy distribution is learned by employing a generative adversarial network architecture. The food energy was estimated from the image based on the generative adversarial network-predicted energy distribution image, and the average estimated energy consumption error was 209 kcal per eating occasion.
AI is transforming diabetes management by enabling early detection of complications, optimizing insulin therapy, and improving patient self-care. Yet, challenges like data quality control, poor technology design, lack of clinical integration, privacy concerns, and imperfection of laws and regulations must be effectively mitigated to increase the adoption of AI in diabetes management in the future.
References and Further Reading
Guan, Z. et al. (2023). Artificial intelligence in diabetes management: advancements, opportunities, and challenges. Cell Reports Medicine. DOI: 10.1016/j.xcrm.2023.101213, https://www.cell.com/cell-reports-medicine/fulltext/S2666-3791(23)00380-4
Makroum, M. A., Adda, M., Bouzouane, A., Ibrahim, H. (2022). Machine Learning and Smart Devices for Diabetes Management: Systematic Review. Sensors, 22(5), 1843. DOI: 10.3390/s22051843, https://www.mdpi.com/1424-8220/22/5/1843
Li, J., Huang, J., Zheng, L., Li, X. (2020). Application of Artificial Intelligence in Diabetes Education and Management: Present Status and Promising Prospect. Frontiers in Public Health, 8, 521222. DOI: 10.3389/fpubh.2020.00173, https://www.frontiersin.org/journals/public-health/articles/10.3389/fpubh.2020.00173/full
Nomura, A., Noguchi, M., Kometani, M., Furukawa, K., Yoneda, T. (2021). Artificial intelligence in current diabetes management and prediction. Current Diabetes Reports, 21(12), 61. DOI: 10.1007/s11892-021-01423-2, https://link.springer.com/article/10.1007/s11892-021-01423-2
Vettoretti, M., Cappon, G., Facchinetti, A., Sparacino, G. (2020). Advanced Diabetes Management Using Artificial Intelligence and Continuous Glucose Monitoring Sensors. Sensors, 20(14), 3870. DOI: 10.3390/s20143870, https://www.mdpi.com/1424-8220/20/14/3870