Managing Chronic Endocrine Disorders With AI

Introduction

Artificial intelligence (AI) is rapidly transforming the field of endocrinology with promising applications in the diagnosis, treatment, and management of endocrine disorders.¹

While AI, machine learning (ML), and deep learning (DL) are often used interchangeably, they represent different concepts. AI refers to the broad field of computer systems designed to perform tasks that typically require human intelligence. Conversely, ML is a subset of AI that enables systems to learn and improve data without explicit programming. DL, a further subset of ML, uses neural networks modeled after the human brain to analyze complex patterns in large datasets.¹

In endocrinology, researchers have actively explored ML and DL algorithms to enhance screening and diagnosis. Their utility has been demonstrated in diabetes care, obesity, and polycystic ovarian syndrome (PCOS), where AI has shown potential in enabling early diagnosis, optimizing treatment, and preventing further complications.¹

Studies show that AI-driven tools can sometimes match or exceed clinician performance, streamlining workflows, conserving resources, and supporting more timely, data-driven clinical decisions. AI thus represents a valuable adjunct to clinical practice in managing complex endocrine conditions.²

In endocrinology, the volume and complexity of digital data are growing at an unprecedented pace. This surge is matched by rapid advancements in computing power, exemplified by the rise of generative AI tools, which are transforming the application of data in clinical practice.³

These developments are enabling the integration of AI into everyday endocrine care, from wearable devices that monitor lifestyle factors to sophisticated systems such as the hybrid closed-loop insulin delivery. As data and computational capabilities expand, so will the potential for AI-driven endocrine diagnostics and treatment innovations.³

According to Leo Celi, MD, associate professor of medicine at the Beth Israel Deaconess Medical Center (BIDMC) and faculty member of the BIDMC Division of Clinical Informatics at Massachusetts Institute of Technology, clinicians must differentiate software as a medical device from non-device clinical decision support.

“And we also have to distinguish wellness products from health products. Wellness products would include wearables that are being used by diabetes patients. But again, wellness products are not under the purview of HIPAA,” he said.

Historical Overview of AI in Health Care

The origins of AI trace back to the early 1950s when Alan Turing introduced what is now known as the “Turing Test.” He posed a profound yet straightforward question: Can a machine think and make decisions like a human? This question laid the groundwork for pivotal developments that would shape the future of AI.⁴

A few years later, the 1956 Dartmouth AI conference, often considered the birthplace of AI, catalyzed broader adoption of the concept and set the pace for its rapid expansion across industries, including health care. By 1964, this momentum had led to tangible innovations, such as ELIZA, the world’s first chatbot, developed by Joseph Weizenbaum at the Massachusetts Institute of Technology (MIT) AI laboratory.⁴

The 1970s saw the rise of AI in medicine with the development of INTERNIST-1, the first artificial medical consultant that helped ease the burden of clinical decision-making and offered physicians a tool to cross-check their differentials using a search algorithm to generate diagnoses from a patient’s symptoms.⁴

The modern era of AI began in the early 2000s, marked by significant advances in health care applications and required refined inputs beyond symptoms. In 2007, IBM launched Watson, which used natural language processing (NLP) and deep question-answering (QA) architecture to extract insights from vast data sources, demonstrating capabilities beyond clinical diagnosis.⁴

Chevon Rariy, MD, board-certified endocrinologist and chief clinical innovation officer at Visana Health, discussed the current AI landscape in managing chronic endocrine conditions. “Over the years, the role of AI in general medicine has evolved from early rule-based expert systems to sophisticated, data-driven platforms that are now reshaping chronic endocrine disease management.”

She continued, “Initially limited by narrow logic and poor scalability, AI’s potential began to expand in the 2000s with the digitization of health records, enabling ML to unlock insights from large, diverse datasets. The 2010s marked a turning point with the rise of predictive algorithms and wearable tech, validating AI’s clinical utility.”

AI in Diabetes Care

Alexander Turchin, MD, MS, a board-certified endocrinologist at Massachusetts General Hospital, also weighed in on the topic. “Within the broader field of endocrinology, one of the most prominent developments has been in diabetes management, particularly among patients on insulin therapy. Increasingly, there is a focus on using algorithm-driven tools to support real-time insulin dose adjustments, allowing for more personalized and responsive treatment.”

The growing burden of diabetes, marked by rising prevalence, complications, and limited specialist access, poses significant challenges to care delivery. Health disparities and poor medication adherence further compound these issues, underscoring the need for more efficient and scalable care models. AI presents a promising solution to identify modifiable risk factors and map out personalized prevention and treatment interventions compared with conventional statistical models.⁵

Recently, researchers have identified the role of AI in the prevention, early diagnosis, and management of diabetes and diabetes-related comorbidities.¹ Scientists have developed generative AI, pattern recognition, and large language models (LLMs) to analyze clinical data, lifestyle factors, and treatment protocols to predict patient outcomes and treatment responses. With these tools, endocrinologists can provide personalized care, adjust medication dosages, and make more informed therapeutic decisions.⁶

One notable study explored the use of a supervised ML model to assess a person’s risk of developing type 2 diabetes (T2D). The team implemented a binary classifier using a shallow neural network architecture trained from scratch to detect non-linear relationships between diabetes onset and various patient health parameters. In their ablation study, the model — optimized using the Adam algorithm — achieved an accuracy of approximately 86% on the test set, with an impressive area under the curve (AUC) of 0.934. These findings highlight the model’s potential as a reliable tool for personalized diabetes risk prediction and clinical decision support.⁷

In an effort to improve precision in T2D care, researchers at Stanford Medicine have developed an AI-powered algorithm that can distinguish among key diabetes subtypes using data from continuous glucose monitors (CGMs). The researchers analyzed CGM data from 54 individuals, 21 people living with prediabetes and 33 healthy individuals, showing that the algorithm could accurately identify metabolic subtypes, such as insulin resistance and beta cell deficiency, with up to 90% accuracy. By capturing nuanced glucose patterns outside traditional clinical settings, the tool offered a more accessible, on-demand method to detect early warning signs and tailor interventions. This innovation could help optimize diabetes management and shape future strategies to prevent diabetes-related complications.⁸

Similarly, promising results reported the use of AI-based algorithms for the diagnosis and classification of diabetic retinopathy (DR). At the core of this advancement is a researcher-developed DL model, which utilized a convolutional neural network (CNN) to assess funduscopic images for DR staging and eye laterality. Using the publicly available Kaggle DR dataset, which included 88,702 color fundus images, the model achieved a sensitivity of 80.28% and a specificity of 92.29% for DR classification. In a separate analysis, the model was trained on 8810 images to detect the laterality of the eye and reached a validation accuracy of 93.28%. Building on these results, the researchers have continued to explore other image-based algorithms applications in endocrinology.¹

Advances are still needed, “but a growing body of evidence suggests that AI tools are shifting diabetes care from reactive to predictive and personalized,” Dr Rariy noted.

“AI tools are transforming diabetes care by enabling proactive, personalized interventions — for example, Medtronic’s Guardian Connect with the Sugar.IQ app uses AI to predict glucose fluctuations hours in advance, helping patients prevent hypoglycemic events,” Dr Rariy said. “Other platforms apply ML to deliver real-time behavioral nudges and flag high-risk patients for timely clinical intervention. These innovations have shown measurable benefits, including lower HbA_1c, improved time-in-range, fewer ER visits, and better adherence,” she added.

AI in Obesity Care

Dr Rariy said AI innovations are helping to move beyond a one-size-fits-all approach, enabling clinicians to offer weight management strategies that are more effective and sustainable.

Recent research also highlights the capability of AI to predict waist circumference using data from a patient’s height, age, weight, ethnicity, and education level. Researchers analyzed data from the Look AHEAD (Action for Health in Diabetes) and National Health and Nutrition Examination Survey (NHANES) trials. They then used a conformal prediction ML method to estimate waist circumference and generate a range of values to validate model certainty.

With 95% accuracy, this ML model outperformed traditional body mass index (BMI) assessment by offering a more precise method to assess obesity-related risk factors. The model generated reliable estimates, making it easier for doctors to predict a patient’s obesity risk without measuring their waist.⁹

Similarly, in the EVIDENT3 (ClinicalTrials.gov Identifier: NCT03175614) trial sponsored by the Biomedical Research Institute of Salamanca (IBSAL), researchers assessed the impact of mobile technology on weight loss interventions. To evaluate the effects of weight loss, body composition, and physical activity among adults with overweight or obesity, the researchers compared an intervention group using a smartphone app and activity tracker wristband vs a counseling-only control group.

At 3 months, the intervention vs control group saw modest improvements, including a greater mean weight loss of 0.84 kg. The intervention group also demonstrated a mean BMI reduction of 0.77 kg/m², mean body fat mass reduction of 1.84 kg, and a mean body fat percentage reduction of 1.22%. By enabling real-time tracking and feedback, AI-enabled wearables and mobile health applications sustained behavioral change and reduced provider burden through targeted alerts that signal clinical intervention.¹⁰

Scientists are also analyzing AI for assisted clinical decision-making across several stages of bariatric care. Study findings have shown that ML algorithms can enhance preoperative risk assessments (difficult intubation in patients with obesity, obstructive sleep apnea [OSA], or pulmonary dysfunction), aid intraoperative pharmacotherapy, and predict postoperative outcomes (mortality, morbidity, weight loss success, obesity-related disease remission, and long-term quality of life). The potential of this technology is particularly relevant given evidence that many physicians hesitate to discuss bariatric surgery with eligible patients.¹¹

AI in PCOS

PCOS is another condition that may benefit from AI, as its early detection is often challenging due to overlapping symptoms with other disorders. Integrating AI into clinical workflows and electronic health records (EHR) could help overcome this challenge by enabling earlier diagnosis, reducing health care burden, and improving patient outcomes.¹²

In a systematic review of 31 observational studies over 25 years, researchers found that ML accurately detected and diagnosed PCOS, with 80% to 90% accuracy in studies using standardized diagnostic criteria. ML demonstrated strong performance across various diagnostic modalities, including ultrasound and lab data, offering a promising tool to address underdiagnosis and misdiagnosis of PCOS.¹²

To address the widespread underdiagnosis of PCOS, which is estimated at 70% globally, researchers developed an automated system for detection and classification using ultrasound images. Their approach involved preprocessing the images with a Gaussian low-pass filter to reduce noise, segmenting key regions using multilevel thresholding, and extracting features with the global image descriptor-multidimensional reduction (GIST-MDR) technique. The processed data were then analyzed using a support vector machine (SVM) classifier, which achieved a high diagnostic accuracy of 96.92%. This model demonstrates the potential of AI-based imaging tools to support early and accurate PCOS diagnosis.¹³

In a study comparing 72 women with PCOS to 73 healthy individuals, researchers applied the BorutaShap feature selection method and random forest (RF) algorithm to identify key predictors in PCOS patients. Out of 58 variables, the most important for PCOS prediction were lipid accumulation product (LAP), abdominal circumference, BMI, and several hormones and metabolic markers. The model achieved 86% accuracy and 97% AUC. Clustering revealed 2 distinct PCOS phenotypes, one characterized by higher BMI, LAP, abdominal circumference, homeostatic model assessment of insulin resistance(HOMA-IR), and insulin levels, suggesting possible metabolic subtypes within PCOS.¹⁴

“Due to its heterogeneous presentation and multifactorial nature, ML models trained on large datasets, including hormone levels, ultrasound findings, metabolic markers, and patient-reported symptoms, could detect PCOS subtypes more accurately than traditional diagnostic criteria alone,” Dr Rariy said. She noted that these models could help clinicians differentiate between phenotypes, such as insulin-resistant vs androgen-dominant, and tailor interventions accordingly.

“This kind of adaptive, data-driven approach can help providers move away from trial-and-error management to truly individualized, whole-person care for women with PCOS,” she added.

Drawbacks and Ethical Concerns

While AI and ML hold significant promises for improving the management of chronic endocrine diseases, they also present important ethical and operational challenges. Key concerns include the potential misuse of AI systems to manipulate quality metrics for financial gain and the risk of perpetuating inherent biases against underrepresented patient populations. These issues could undermine efforts to deliver equitable, high-quality care.²

A further drawback lies in the potential erosion of the physician-patient relationship. Although AI can assist with diagnostic, therapeutic, and even procedural tasks, these technologies cannot replicate the emotional support, contextual understanding, and ethical decision-making that physicians provide. Maintaining a human-in-the-loop approach is essential to ensure that medical professionals remain central to patient care, preserving the trust and compassion foundational to effective clinical practice.¹

Moreover, integrating AI into health care raises significant security and privacy concerns, particularly regarding managing sensitive patient data. On the other hand, safeguarding health information will maintain public trust and ensure ethical compliance. As the field evolves, the responsible application of AI will prioritize transparency, human oversight, and the continued delivery of compassionate care.²

One of the biggest challenges in AI development for health care, Dr Celi stated, is the reliance on EHRs as foundational data. “EHRs were originally designed for administrative tasks — billing, documentation, and quality tracking — not for capturing the complete clinical picture needed for accurate AI training. As a result, these records often lack the depth and precision required to understand authentic relationships between clinical features and outcomes,” he said.

“This leads to models learning inaccurate or misleading associations which can compromise patient care. Without improving our understanding and curating data before building AI models, we risk serious flaws in AI-driven decisions, especially in complex conditions like diabetes, PCOS, COPD, and heart failure,” Dr Celi added.

Looking Ahead

Despite their potential to lead AI/ML integration in health care, endocrinologists often lack formal training in these technologies. Most endocrine subspecialty programs do not equip trainees with the skills or confidence to use AI for diagnostics.²

To better prepare physicians for the AI era, experts recommend that the medical community should focus on:

• Shifting curricula from passive information acquisition to active knowledge management and communication;
• Training physicians to work alongside AI and ML tools in clinical workflows;
• Emphasizing skills to assess and apply AI-generated data in critical clinical decision-making; and,
• Reinforcing empathy and compassion in an increasingly tech-assisted health care environment.²

Looking ahead, Dr Rariy said AI will play an increasingly transformative role at the intersection of endocrinology and women’s health, where chronic conditions such as PCOS, obesity, diabetes, and thyroid disorders often coexist and evolve across a woman’s lifespan.

“AI will enable a more interconnected, systems-based approach, analyzing patterns across hormonal cycles, metabolic data, mental health indicators, and social determinants to detect risks earlier and tailor interventions with greater precision,” she added.

Adding to these comments, Dr Turchin stated that the evolving role of AI in clinical care is best viewed as supportive rather than substitutive.

“In practice, AI tools such as those offering diagnostic suggestions or treatment options serve as aids to the clinician, who ultimately partners with the patient to make the final care decisions. For instance, in glucose management, algorithms can provide valuable recommendations,” he noted.

“However, the process always begins with a discussion: we present patients with various device options, outline the pros and cons, and empower them to choose based on their preferences and needs. Even after adoption, patients retain full control, they can override the device’s suggestions and consult their care team or the manufacturer’s support team for guidance,” Dr Turchin explained.

“AI is here to enhance and not replace clinical judgment and patient-centered care. While these tools can significantly improve efficiency and personalization, human clinicians remain central to interpreting, validating, and contextualizing AI-driven insights,” he concluded.

This article originally appeared on Endocrinology Advisor