Role of Artificial Intelligence in Diabetes Mellitus Care: A SWOT Analysis

Abstract

Diabetes mellitus has become one of the major public health problems in India. Chronic nature and the rising epidemic of diabetes have adverse consequences on India’s economy and health status. Recently, machine learning (ML) methods are becoming popular in the healthcare sector. Human medicine is a complex field, and it cannot be solely handled by algorithms, especially diabetes, which is a lifelong multisystem disorder. But ML methods have certain attributes which can make a physician’s job easier and can also be helpful in health system management. This article covers multiple dimensions of using artificial intelligence (AI) for diabetes care under the headings Strengths, Weaknesses, Opportunities, and Threats (SWOT), specifically for the Indian healthcare system with a few examples of the latest studies in India. We briefly discuss the scope of using AI for diabetes care in rural India, followed by recommendations. Identifying the potential and challenges with respect to AI use in diabetes care is a fundamental step to improve the management of disease with best possible use of technology.

INTRODUCTION

Diabetes mellitus is a serious, long-term condition that occurs when raised levels of blood glucose occur because the body cannot produce any or enough of the hormone insulin or cannot effectively use the insulin it produces.[1] About 422 million people worldwide have diabetes, the majority living in low- and middle-income countries, and 1.5 million deaths are directly attributed to diabetes each year.[2] By 2045, International Diabetes Federation projections show that 1 in 8 adults, approximately 783 million, will be living with diabetes.[1] Furthermore, diabetes is the 14^th largest cause of Disability Adjusted Life Years (DALYs) worldwide.[3] Chronic nature and the rising epidemic of diabetes have everlasting consequences on the nation’s economy and health status.[4]

There has been a huge impact of technology in the medical world with the recent advancements. Health outcomes may be adversely affected in a matter of seconds for individuals who may not be able to reach a hospital or receive emergency treatment. Technology bridges this gap in distance, time, and other resources for all people to whom its benefits are extended.[5] In the last few years, artificial intelligence (AI) has become a useful tool in the medical field for various purposes. Nowadays, machine learning (ML) algorithms, which are in fact applications of AI, are used for automatic analysis of high-dimensional biomedical data. Diagnosis of liver disease, skin lesions, cancer classification, risk assessment for cardiovascular disease, and analysis of genetic and genomic data are some of the examples of biomedical application of ML.[6,7]

Artificial intelligence

The British mathematician Alan Turing (1950) was one of the founders of modern computer science and AI. Modern medicine is faced with the challenge of acquiring, analysing, and applying a large amount of knowledge necessary to solve complex clinical problems.[8] The past 10 years have seen remarkable growth and acceptance of AI in a variety of domains and in particular by healthcare professionals. AI-enabled clinical decision-support systems may reduce diagnostic errors, augment intelligence to support decision making, and assist clinicians with EHR (Electronic Health Record) data extraction and documentation tasks.[9] The ultimate aim of AI is to make things fast and easy as well as accurate. AI methods simulate human intelligence and perform several tasks like classification, visual and auditory perception, and decision making.

Machine learning

The term ML comes under a broader heading of AI. ML uses statistical techniques to give computers the capability to ‘learn’ particular assignments without being explicitly programmed through algorithms that are intended to copy human intelligence utilising knowledge from the surrounding environment.[10,11] ML methods identify and learn patterns in high-dimensional data to create prediction and classification models. Using huge data sets algorithms are designed through some logic, probability, statistics, control theory, and so on to analyse the data and retrieve the knowledge from the past experiences.[12]

ML has become a useful tool in the expanding field of bioinformatics. It has been effectively used in the early prediction of cardiovascular diseases,[13] predicting survival rates in cancer patients,[14] stroke predictions,[15] and management of cancer,[16] and work is being done in predicting the onset of Alzheimer’s and Parkinson’s disease also. ML methods are now increasingly applied to the clinical epidemiology of diabetes, particularly to help risk stratification decisions, decisions about the changing or escalation of therapy, or diagnoses and decisions based on complex input data such as images or continuous glucose monitors.[17]

Machine learning methods

There are mainly three types of ML: 1) supervised learning (input data are tagged), 2) unsupervised learning (input data are not tagged), and 3) reinforcement learning (technique in which the computer learns by trial and error and by collecting overall rewards imitating natural intelligence). ML can be used for clustering, classification, association, and prediction regarding diabetes and its risk factors by using certain algorithms. A few commonly used algorithms in the area of diabetes mellitus are regression, decision tree,[17,18,19,20] naive Bayes,[21,22] neural networks,[17,20,23,24] ANN (artificial neural networks)/SNN (simulated neural networks), support vector machine,[22,23,25,26,27] ensemble machine learning,[20,23,27] and deep learning.[28,29]

Table 1 shows different AI approaches in diabetes management in India in recent years. The first column included the author’s name and year of publication; the second column includes the details – aspects of diabetes care for which AI has been applied, short description, and ML methods used in that study. The third column included remarks regarding how in future the study can be used or extended to improve our knowledge. It is not an exhaustive list. We have tried to include different studies to cover several aspects of diabetes care like diagnosis, screening, and drug selection.

Table 1.

Application of AI for diabetes care in India (some recent studies from India)

Author and year	Applications in Diabetes Care	Remarks and future scope
Natarajan et al.,[30] 2019	Diagnostic accuracy  This study compares the diagnostic accuracy of a smartphone-based AI system with ophthalmologist judgement in patients with referable diabetic retinopathy or any diabetic retinopathy in Mumbai, India.  ML technique: Neural Network	The authors found it as a cost effective method. Assessment of such an AI system is lacking.
Birk et al.[31] 2021	Screening for pre diabetes in rural India  The authors developed a predictive tool to screen for prediabetes using survey data from an FFQ (Food Frequency Questionnaire) to compute the Global Diet Quality Score (GDQS). They developed a prediction algorithm that utilizes measures of diet quality and other predictors of diabetes risk (age and tobacco use)  ML Technique used: Random guessing model, Logistic regression model, Least absolute shrinkage and selection operator regression, Elastic net, Random forest, Generalized linear mixed-effects model	The main objective behind this study was to prioritise individuals as high risk and who need further laboratory tests (instead screening all individuals). The authors recommended future studies to examine the utility of the GDQS in screening for other non-communicable diseases.
Singla et al.,[32] 2022	Prediction of diabetes drugs  This study used input as patient variables to predict all diabetes drug classes to be prescribed. The aim was to develop a ML technique for prescription of anti-diabetes drugs when provided with appropriate clinical data.  ML Technique used: Decision tree, Random forest	Needs to be validated across multiple practices for wider application. Currently, predictions are at the drug class level; predictions at the drug dosage level would make it more useful in clinical practice.
Kulkarni et al.,[33] 2022	Early prediction of diabetes  The authors developed a methodology using ML Algorithms for Diabetes Disease Risk Prediction in North Kashmir. ML technique used: component-wise boosting, multilayer perceptron, probabilistic neural network, random forest and extreme gradient boosting (XGBoost) algorithm.	Needs validation. This algorithm can help in early detection of diabetes and pre-diabetes after robust validation in external datasets.
Nanda et al.,[34] 2022	Diabetic foot ulcer (Example of use of AI as classifier)  Evaluation of risk factors of diabetic foot ulcers and its severity.  ML technique used: Support vector machines (SVM-Poly K), Naive Bayes (NB), K-nearest neighbour (KNN), random forest (RF) and three ensemble learners.	Authors mentioned that in future this work can be extended for development of a “white box”-based prediction system giving insights into the process of classification
Menon et al.,[35] 2023	Patient tracking  This research suggested a system for diabetic patient tracking based on ML. The proposed technique can be used to assess the person’s dependence, forecast his future health status, and foresee its decline before potential consequences. (example of AI based E-health)  ML technique used: ASV-RF (advanced-spatial-vector-based Random Forest)	In terms of accuracy, the suggested strategy worked superior to other current approaches. In the future, the proposed framework can be extended with various large datasets.

Numerous articles from different countries have been published on use of ML algorithms in diabetes care. However, literature from India is relatively less because we are still in the beginning phase of using AI in our healthcare system. In this article, we have specifically discussed the aspects of application of ML methods for diabetes care in India under the headings of Strengths, Weaknesses, Opportunities, and Threats (SWOT). This shall provide some insight into the potential and challenges that may be encountered while using such methods in diabetes management.

Strengths

Non-invasive: ML methods are non-invasive. Specifically in diabetes, patients often resist repeated finger prick tests for screening and diagnosis, and it is uncomfortable as well as in the regular glucose monitoring due to the discomfort caused in the process. AI, to a certain extent, can help in identifying the risk of diabetes in a non-invasive way, like use of tongue images (color and texture) in establishing a non-invasive diabetes risk prediction model.[36] The compliance and acceptance towards non-invasive methods for diabetes screening are likely to be more as compared to a finger prick, which would lead to improvement in identifying undiagnosed cases of diabetes, which is not a small number in India. However, the accuracy of such methods must be considered before a physician can rely on such ML approaches in screening.

Detail-oriented and fast: A computer reads any provided data in a much more detailed manner than humans. The main advantage of letting a machine read the data is that a machine can identify patterns and details that humans cannot with similar speed and accuracy, especially capturing the details in images (e.g., fundus images), x-rays, audio, and so on. Hues and shades of colors in images and pitch of sounds in audio can be identified in a precise manner by a computer. In addition, once properly trained, ML methods are less time-consuming. In healthcare facilities where patient load is huge, ML methods can make patient management is easy to some extent. While running algorithms, several factors can be considered simultaneously. Furthermore, AI software is not encumbered by human issues such as fatigue or other environmental interruptions that may slow it down or reduce its accuracy with time.[37]

Application in telemedicine: In remote areas where the physical presence of a specialist is difficult, ML algorithms can help in not only prediction and diagnosis of diabetes but also screening for its complications and follow-up of patients. A reliable ML model which can be used by trained staff or trained community health workers can help in such scenarios. Once identified, such patients can be referred to healthcare facilities. Globally, telemedicine programs have demonstrated accuracy in classification of DR (diabetic retinopathy) into referable disease and into stages with accuracy in a cost-effective manner and with sufficient patient satisfaction.[38]

Relatively inexpensive: ML is more economical than machine programming. After a model is trained, it takes nearly no cost for prediction and other applications. Several platforms can be utilised for ML, like Kaggle, Anaconda, and Tensor flow. These can be used for practice by health professionals. However, cost-effectiveness analysis studies on AI can help to gain more insight into it.

Weaknesses

Inability to handle new scenarios and complexities: An ML algorithm gives output based only on previously provided data. Diabetes is a complex disease that affects multiple systems. Physicians face different clinical complexities in patients, and not every scenario can be handled by an algorithm. Several times, diabetes may result in unpredicted events in a patient. Trained algorithms would not be able to assess any new inputs. In those cases, a doctor shall be required to make decisions based on his expertise. This is one of the important reasons why AI cannot replace doctors specifically for diabetes care.

Requirement of Big data sets: Though ML techniques are very advanced, their learning requires big data sets. Maintaining big data with details of real diabetes patients requires a robust EHR system. Electronic record keeping is not a well-established area in many countries including India. Big data of real patients need to be provided to the machine to get trained, to have transparency, and to form a reliable model for use. Fabricated data or extrapolated data shall not be reliable. Datasets used for training are often incomplete and noisy and may have inherent biases.

Issues with generalisability: By ‘Generalisability’, we mean application of a particular ML algorithm in new settings (where it was not originally generated). Current algorithms for diabetes diagnosis have some limitations and are not tested on different datasets or people from different countries, which limits the practical use of prediction methods. Diabetes epidemiology varies globally, and it keeps on changing with time. Several genetic, racial, and ethnic factors also play significant roles in diabetes overall. ML algorithms developed in one institution, state, or country may not be accurate for others. The models might not be generalisable when applied to datasets from other health care centers or geographic locations, rendering the ML model useless for a broad application.[39]

Behavioral factors: Every diabetes patient requires a plan for day-to-day life to maintain the glycemic control. If the patient fails to make the behavioural adjustment necessary, for example, losing weight, scheduling a follow-up visit, filling prescriptions, or complying with a treatment plan, these plans go unsuccessful. Human behavioural factors shall always be a concerning factor in chronic diseases like diabetes that require motivation to stick to a particular lifestyle. Issues like non-adherence and non-compliance cannot be completely resolved by AI alone. Mobile apps can partially help the patient, but ultimately, it is the patient’s motivation to adhere to long-term treatments.

Accuracy and reliability: ML methods are not fully accurate. There are several studies that measure the accuracy of their developed model. While taking decisions regarding any diabetes patient, it is necessary for healthcare workers and patients to consider the fact that the ML methods are not fully reliable. This may fix accountability issues for doctors. Any decision should not be solely based on AI outcome.

Opportunities

ML methods hold the potential to improve diabetes care in India to a great extent.

AI-assisted diabetes care: There are several dimensions of diabetes management. Diabetes is the most popular chronic clinical condition targeted by mHealth. There are several mobile apps for diabetes patients that are helpful in providing guidance in lifestyle modifications, foot care, and diet. Several counseling mobile apps function through the application of chat boxes. Patients who are less mobile can benefit through it the most. In recent times, the management of diabetes is becoming more patient-specific. While making decisions, the occupation, lifestyle, dietary habits and preferences, season, residential area, and so on should be explored to provide the best treatment, lifestyle changes, diet charts, and follow-up schedules to practice precision medicine. Though humans can consider a few factors simultaneously while making decisions, a machine can consider numerous factors simultaneously in almost no time. Algorithms like decision trees and decision forests can help in deciding and prescribing individual specific management faster. Also, every patient responds to medication differently. By providing data of drug-related profiles, prediction of side effects can be made by using ML methods. Accordingly, the most appropriate drug can be prescribed.

Preventive medicine: The risk prediction ability of ML methods can help in planning several diabetes prevention strategies in a more focused manner. ML can predict the risk of developing diabetes in early ages and also the probability of developing complications. At the community level, these methods can be utilised to screen patients for diabetes and bring them to healthcare facilities. It is estimated[40] that 98% of vision loss from DR and macular oedema is avoidable through improved prediction, early detection, and treatment strategies, and its cost-effectiveness is well established.

Improving India’s healthcare system: Shortage of skilled workforce can be dealt by using AI methods. ML methods can be used at several places where the doctor–patient ratio is not satisfactory. The Government should invest in training doctors and community healthcare workers to make them familiar with the common ML techniques that can be used in healthcare. Specifically, diabetes is increasing in India at an alarming speed. There should be a separate portion of healthcare budget to be invested in AI and ML domains.

Strengthening EHR (Electronic Health Record) of diabetes and overall: Huge data are required for a model to develop. Therefore, if a country is planning to embed ML in their healthcare system, the fundamental step should be emphasised on EHR maintenance. This shall not only be helpful in future for ML techniques but also help in maintaining/surveillance/assessing current pictures on a particular health condition. For example, a diversity of images is required for diabetic retinopathy ML platforms. Quality of images (e.g., fundus) data can be improved by using the best available technology cameras. By providing the best quality fundus images, more and more details can be identified by machine and it becomes more detail oriented and accurate.

Improving the multidisciplinary approach in healthcare: Continuous communication between data scientists and medical staff can reduce research-related problems. AI integration in the Indian healthcare system gives the opportunity to improve the multi-disciplinary setup, which is a fundamental requirement for AI to be used efficiently for healthcare. Doctors, ML experts, and biostatisticians need to work in collaboration for its best utilisation.

Threats

There are several issues with the use of AI that may limit its use in diabetes care.

Lack of empathy: Empathy and compassion is a fundamental need for patient-centred care. AI cannot replace human empathy. Not just medicines but compassion, kindness, and empathy are equally important for a patient’s management. One of the biggest threats for using AI in healthcare is that the human-to-human interaction may become limited. Chronic diseases like diabetes require regular and life-long mental support to improve patient satisfaction and drug compliance and therefore the overall management. Meeting the physician on a regular basis and sharing the experiences and problems directly to the doctor have a huge psychological impact to keep the patient motivated. Talking to caregivers also gives them strength to be involved in a patient’s management. Due to such reasons, patients will most of the time want to talk to a doctor and not to an ML platform or a chatbot or some virtual assistant. Non-involvement and lack of trust of patients in ML methods shall be an issue, especially in the initial stages of implementation.

Non-authentic, low-quality data, and other data-related issues: Wrong or fake data given to the machine shall not be of much use. To ensure that the data are real patient data, we shall need some identifier of a patient which might have ethical problems and privacy problems. Since there might be racial differences, prediction models to be accurate shall require genetic and race-specific data. Data from other countries cannot be used to predict the disease in India. In addition, quality of data might also affect the overall accuracy of the ML model. For example, bad-quality images of fundus in case of DR can also misclassify the patients and may lead to wrong decision making. These issues may limit the use of AI for diabetes. Also, India is the diabetic capital of the world. If we make a good EHR system, our data can be used globally throughout the world to develop trained algorithms. However, data must be anonymised before they are available for use in machine learning. There must be strict measures that ensure unauthorised access to maintain confidentiality of patients’ data.

Communication gaps: In remote areas where Internet connection is low or absent, it may be challenging to practice telemedicine. Diabetes is a disease that requires attention from different specialities, like endocrinologists, ophthalmologists, and nephrologists. Since use of AI in healthcare requires a robust transparent communication system between doctors of various specialities, miscommunication might lead to wrong decision making. Not only hospital doctors but also certain field workers should also be in good communication with doctors.

Poor response of doctors and patients: Humans have a nature of resistance to change. Many doctors may not prefer to use ML methods because of not being well versed with AI platforms and their applications in healthcare. Several training sessions would be required for doctors who focus on the use of ML methods focused on healthcare. Since it is a new area, not only patients but also doctors may not be confident to rely on ML methods. This becomes more significant in life-long chronic diseases like diabetes where the condition of the patient may fluctuate from good to bad in a small time period.

Ethical and legal issues: ML models should not be in the wrong inexperienced hands. There should be regulations on the use of ML methods so that only professionals can use them. Such methods if used by quacks may lead to malpractice. Not only that, it might cause mismanagement of diabetes and its complications that shall ultimately increase the burden and severity of disease and hospital admissions. Regulation, legal causes of action such as medical malpractice and product liability, intellectual property, and patient privacy all have real implications for the way AI is developed and deployed.[41] Furthermore, a patient’s autonomy must always be maintained while practicing AI-assisted diabetes care.

Black box nature of some algorithms: The decisions of many ML models are not easy to understand and interpret, especially with increased complexity of the input data. It can become difficult to understand and justify how a specific algorithm came to a certain output. This raises legal and ethical concerns. Research on increasing the interpretability and thereby decreasing the “black box” nature[33] of ML models will play a crucial role in the acceptance of such models within clinical decision making.

Accountability: An interesting consideration is the responsibility for error induced by the ML model as physicians can be held legally responsible for their decisions, but it is unclear who can or should be held responsible when an ML model makes the wrong decision.[33] It may be difficult to establish accountability for the mistakes made by ML algorithms. It may further increase legal and ethical concerns.

Scope for using AI for diabetes care in rural India

AI can be a great contributor to achieve Universal Health Coverage. Rural healthcare in India is still suffering from lack of doctors; specifically specialists. However, diabetes is becoming equally prevalent in rural areas, and slowly, it is increasing in hilly and tribal areas as well. AI can help in identifying undiagnosed cases of diabetes in such areas. Following are some aspects where AI can be used in rural health in our opinion:

1. Telemedicine has been practiced in the past in India and has been regarded as a successful initiative in rural areas. The Chunampet Rural Diabetes Prevention Project, by V Mohan et al.,[42] was a successful model for screening and for delivery of diabetes health care and prevention to underserved rural areas in developing countries such as India. More such initiatives should be practiced to develop a good foundation for practicing telemedicine in future on a larger scale.

2. As the government of India has upgraded sub-centers into Health and Wellness Centers (HWCs), further upgradation can be done by implementing AI-assisted primary care. Several early prediction and non-invasive diagnostic techniques have been proposed for diabetes. Implementing such techniques can be helpful in managing the disease in early stages only.

3. For diabetes patients, in rural and remote areas where a professional dietician is not available, an accurate AI algorithm can help in a patient specific dietary chart, depending on habits, personal preferences, and liking of the patient. This may not be possible at hospital levels for each and every patient because of patient load. It can be practiced in HWCs.

4. In addition, as diabetes requires regular screening for complications, this screening can be done at primary levels by utilising trained paramedical staff. In such cases, a trained AI device handler can be appointed if a physician is not available. It will be convenient to the patient and the health system. Patients can easily approach nearby health facilities and get themselves screened instead of going to a tertiary care hospital. In addition, it will decrease the patient burden in a tertiary care hospital. Only those patients who need to be referred for retinopathy shall be going to tertiary hospitals. This shall improve the screening and early detection of diabetes complications. Recently, several online AI platforms, including google, have also worked on screening for DR.[43]

5. Another application can be community based AI-assisted screening of diabetes by going house to house, specially using non-invasive methods. Various non-invasive[36,44,45,46] approaches have been proposed, like screening with tongue images and blood flow properties. People will be more receptive to non-invasive methods. It shall increase screening in the population. Field workers can be provided with certain devices which can tell the signs of complications. However, for such applications, we need to provide simpler devices which can be handled by community health workers. Guo et al.[47] have proposed a multilevel medical AI service network, including frontline medical AI system (basic level), regional medical AI support centres (middle levels), and a national medical AI development centre (top level). It might take some time for India to develop such a system; however, it shall be beneficial.

6. Training of para-medical staff and community health workers to use AI devices is necessary. ML methods have to be successfully implemented. One such example is a study[48] done for capacity building for diabetic retinopathy screening by optometrists in India, where they trained optometrists for a 7-month long fellowship in DR and mentored participation as co-facilitators in 1-day orientation workshops on DR screening guidelines across India. Feasibility of such methods need to be explored more by doing implementation research.

RECOMMENDATIONS AND CONCLUSION

This article has covered multiple aspects of using ML methods for diabetes under the healthcare system of India. National Programme for Prevention of Noncommunicable Diseases (NP-NCD) along with National Digital Health Mission (NDHM) can be a platform to integrate ML methods in management of diabetes, particularly in the government sector.

Some recommendations for the limitations mentioned earlier: Basics of ML should be taught to healthcare workers with regular training sessions that focus on the use of algorithms appropriate for diabetes and overall healthcare. By investing in and improving the healthcare infrastructure, several technical issues can be addressed, like data storage, communication, and Internet issues. A disciplined EHR system, if achieved, can improve data quality. To address generalisability issues, wide spectrum, real, good-quality data can help to some extent. Further, every country and state and maybe every health facility will have to make their own algorithms. We recommend use of ML algorithms in situations where doctor–patient communication is not much needed or is required to a limited extent for diabetes patients. It can be used in imaging (for DR), screening (early diagnosis), patient load management, smooth operations of OPDs, predictions, and patient-specific approaches (dietary chart algorithms, etc.). In clinical management of complex scenarios, it is better that a patient interacts with a real doctor. So, the Human-In-Loop ML approach seems to be the best for AI-assisted diabetes care. Particularly, at the decision making stage, it should be a human who is finally deciding for the patient.

A number of studies are being published regarding development of an AI framework using ML methods in diabetes management along with comparative studies between different AI techniques. The authors believe that side by side, there is a need to take a step further to try to apply these proposed frameworks on the actual population (maybe on small samples) to see the feasibility and to identify practical problems regarding implementation of AI methods for diabetes care in the Indian healthcare system.

In conclusion, the healthcare system should identify, differentiate, and prioritise the areas where it shall be relatively more appropriate and convenient to use ML methods for diabetes care (by considering cost-effectiveness). Finally, the popular opinion that AI might replace doctors in healthcare is quite unlikely to happen. The clinical complexities of diabetes cannot be handled by ML algorithms solely. Instead, AI can act more like an assistant to the physicians making their job easier and faster.

Authors’ contribution

PK and SVM have contributed to the concept and overall intellectual content of this article. Literature review, manuscript writing and editing was done by PK. Manuscript review and supervision was done by SVM and MKU.

Conflicts of interest

There are no conflicts of interest.

Funding Statement

Nil.