Transforming the cardiometabolic disease landscape: Multimodal AI-powered approaches in prevention and management

SUMMARY

The rise of artificial intelligence (AI) has revolutionized various scientific fields, particularly in medicine, where it has enabled the modeling of complex relationships from massive datasets. Initially, AI algorithms focused on improved interpretation of diagnostic studies such as chest X-rays and electrocardiograms in addition to predicting patient outcomes and future disease onset. However, AI has evolved with the introduction of transformer models, allowing analysis of the diverse, multimodal data sources existing in medicine today. Multimodal AI holds great promise in more accurate disease risk assessment and stratification as well as optimizing the key driving factors in cardiometabolic disease: blood pressure, sleep, stress, glucose control, weight, nutrition, and physical activity. In this article we outline the current state of medical AI in cardiometabolic disease, highlighting the potential of multimodal AI to augment personalized prevention and treatment strategies in cardiometabolic disease.

INTRODUCTION

The ability to model complex relationships from massive datasets and derive connections between diverse inputs and outputs has catapulted artificial intelligence (AI) into the mainstream of science. The impact in biomedicine has the potential to be transformational.^1–3 This is especially true as the layers of human data grow deeper and more longitudinal.

The practice of medicine has traditionally involved a medical provider using a limited range of patient-level data inputs such as the medical history, current concerns or symptoms, a physical exam, and possibly a few dedicated labs or imaging studies to map out prevention and treatment goals based on best practices. Conventionally, linear models were used to provide improved risk prediction and prognosis based on these data, but machine learning has taken this to a new level. Broadly speaking, machine learning is a branch of AI that aims to establish and recognize patterns between specific inputs, here a patient’s medical data, and outputs, such as future disease risk, response to therapy, or prognosis.^2,4–6 AI has transitioned from supervised learning where labeled data and ground truths guided model development to a point now where the data can learn from itself.⁷ As the type and content of medical data grew and diversified, the ability of human experts to review, verify, and classify data has been outpaced. This is where the current phase of AI comes in with transformer models and generative AI that allows for the incorporation of multimodal datasets for self-supervised learning and prediction.

In cardiology, machine learning quickly took hold, incorporating two of the most common diagnostic tests in use today: the electrocardiogram (ECG) and echocardiogram. Early algorithms focused on automated ECG and echocardiogram interpretation.^8–13 Soon after, these algorithms were trained on large-scale, longitudinal datasets to predict adverse outcomes, such as mortality, including disease onset such as future atrial fibrillation or left ventricular dysfunction.^14–18 Recently, a random forest model of ECG diagnosed angiography-proven coronary occlusion in the setting of heart attack, markedly improved over clinician diagnosis.¹⁹ In this way, AI provided more accurate diagnosis than traditional metrics and insight into the presence of disease in an asymptomatic setting, or predicted high risk of future disease, but was oftentimes limited to algorithms trained on a single modality. However, machine learning modeling had to adapt to the increased number and complexity of data sources. This was accelerated by evolution of transformer models (Figure 1) that moved away from supervised, step-by-step, sequence-based models and have the ability to analyze unlabeled datasets in a parallel fashion without the restraints of needing to process the data in a specific order.^20,21 Transformers increased the amount and types of data analyzed and speed of processing, leading to an explosion of generative AI models such as ChatGPT and GPT4. Details on novel methods for fusion and multimodal machine learning modeling were reviewed in detail elsewhere.^22–24

Figure 1. Comparison of traditional RNNs and transformer models in sequence processing.

Traditional recurrent neural networks use local sequential processing, as illustrated by a stepwise approach from word to word, which limits scalability and adaptability, whereas transformer models have self-attention allowing a focus on the interaction of individual positions in a specific context so as to build long-range relationships. Created with BioRender.com.

For patients, there are multiple data sources taken into consideration. These come in the form of not only vital signs (blood pressure, heart rate, oxygen saturation, and temperature); laboratory values; medications; imaging data from X-rays, MRIs, CTs, and ultrasound; pathology reports; and medical histories that can be obtained from the electronic health record (EHR) but also specific symptoms, sleep, and physical activity data from wearable sensors and genome sequence results, all which would ideally be considered in real time to determine the next logical steps in prevention or treatment strategies. As physicians, we have created checklists and heuristics to enable us to evaluate these various data streams and interpret them the best we can, though we continue to fall short. This is the point where the true potential of multimodal medical AI becomes apparent.²⁵

Here, we focus our attention on the areas where multimodal AI will likely have the greatest impact on the reduction of cardiometabolic disease (Figure 2). We highlight existing work in this space and discuss the opportunities to integrate multiple data streams for personalized prevention.

Figure 2. AI-driven integration of multimodal data for personalized cardiometabolic disease management.

Integrating diverse data layers, including digital sensor technologies, health records from hospitals and clinics, laboratory findings, and imaging data, as well as innovative genomic and biomarker inputs through artificial intelligence algorithms, has shown promising outcomes. These efforts aim to decrease cardiometabolic disease by enabling a personalized approach to optimizing risk factors and managing the disease. Created with BioRender.com.

RISK ASSESSMENT FOR TARGETED PREVENTION

Clinical risk tools

The pathophysiology of coronary artery disease (CAD) is complex and involves the combinatorial interplay of factors influenced by an individual’s personal genetics, lifestyle, and environment. CAD is marked by atherosclerotic plaque formation in the main epicardial arteries supplying oxygen and nutrients to the heart muscle. Plaque is made up of lipid-rich particles, smooth muscle, calcium, and immune cells and typically builds up over time, starting in early adulthood and manifesting as frank disease in mid to late life. Some plaques are unstable and can rupture, leading to thrombosis and resulting in partial or total occlusion, clinically manifest as an acute coronary syndrome.

Large-scale population studies have identified the most common modifiable factors for CAD, including blood pressure, healthy diet, physical activity and exercise, obesity and weight management, diabetes and glucose regulation, tobacco use, cholesterol levels, and sleep.^26,27 Many of these clinical risk factors were integrated into risk assessment calculators, such as the atherosclerotic cardiovascular disease (ASCVD) risk calculator and QRISK, to help identify individuals at an elevated risk for developing CAD for targeting risk factor optimization strategies including statin therapy.^26,28 These tools remain the standard of care in risk assessment, though they are known to over- and underestimate risk in various populations.^29,30

Machine learning models improve on traditional risk assessment tools and evaluate their performance in broader, more ethnically diverse populations. In an assessment of 434,604 participants of the UK Biobank, a machine learning-based model incorporating 473 variables showed a modest improvement over the Framingham score, AUC 0.774 and 0.724, respectively.³¹ As such, the machine learning-based model predicted cardiovascular disease more accurately in 368 more cases, out of a total 4,801, over Framingham, over a period of 5 years. Using EHR-derived data in a diverse patient population (49% non-Hispanic white, 30% Asian, and 8% Hispanic), another group showed AUC of 0.835 for predicting ASCVD using a model of 1,175 EHR variables compared with AUC 0.775 when using the traditional clinical risk tool (pooled cohort equations).³²

Polygenic risk scores

Although the heritable nature of CAD had long been appreciated from family and twin studies, the advances in genome sequencing and large-scale genome-wide association studies (GWASs) illuminated a multitude of individual variants associated with disease.^33,34 To quantitate how these many individual variants ultimately influenced a person’s overall genetic risk for CAD, polygenic risk scores (PRSs) were developed and shown to be effective for identifying high-risk persons.^35,36 It is important to note that individuals with high CAD PRSs are not necessarily destined to disease and healthy lifestyle attributes have profound effects on reducing their actual risk overall, thus the emphasis of using CAD PRSs in primary risk assessment.³⁷ In one study of younger individuals (<50 years old), those in the highest genetic risk (top 20%) had three to four times greater risk of heart attack than those with the lowest genetic risk.³⁸ Additionally, there was added benefit of combining CAD PRS with ASCVD risk calculations in this high genetic risk population for predicting future events: AUC 0.74 compared with 0.53 for clinical risk calculation alone. Other efforts have shown similar modest improvements of risk calculation when adding clinical risk to PRS, which help to reclassify individuals more accurately to either higher or lower risk tiers.³⁹

Initial efforts to utilize AI models in the PRS space but only on genetic information did not show improvement over PRS by logistic regression.⁴⁰ However, multimodal AI incorporating CAD PRS with 51 of 13,782 possible predictors (obtained from survey data, biomarkers, body measurements, and clinical data) did improve event discrimination compared with traditional clinical risk calculators.⁴¹ Even when incorporating only 29 clinical variables with CAD PRS, improvements in risk discrimination over QRISK, Framingham, and PCE using a neural-network-based model were apparent, especially, as was seen above, in individuals younger than 50 years of age.⁴² Models using longitudinal clinical data from EHR were superior to ASCVD risk calculation with additional improvement adding 204 SNPs.⁴³ In an impressive EHR-cohort study of 95,935 participants, machine learning models not only achieved an AUC of 0.95 for CAD risk prediction, but their model accurately quantified the degree of coronary stenosis, complexity of disease, and risk of death.⁴⁴ In a cohort of 1,329 patients, unsupervised machine learning using genetic and 155 phenotypic variables successfully grouped patients with atherosclerosis into distinct subsets, each having unique trajectories (mortality, cerebrovascular disease, and heart attack).⁴⁵ While the traditional ASCVD risk calculator is not clinically sophisticated and allows for widespread use in primary care, as we continue to develop deeper data for patients including genetics, the benefits of multimodal AI to assess risk and assist in clinical optimization strategies, including medications, are clear. Interestingly, there appears to be a relationship of medication purchases, changes, and even treatment discontinuation with PRSs for CAD, diabetes, and hypertension.⁴⁶ As larger, multiethnic studies are conducted, we continue to discover new variants associated with CAD and gain insight into the interplay of CAD with comorbid diseases such as obesity, diabetes, and nonalcoholic fatty liver disease.^47–49

Advanced imaging for CAD risk stratification and diagnosis

While EHR-derived and genetically informed risk assessment platforms are widely used to date, research using additional data streams has also shown to be powerful for CAD risk assessment and diagnosis. A machine learning algorithm featuring lifestyle questionnaires, clinical blood results, and 24-h ambulatory heart rate and blood pressure monitoring in an Asian population outperformed FRS for association with elevated coronary artery calcium (CAC), a marker of subclinical atherosclerosis and future risk.⁵⁰ Clinical imaging such as carotid artery ultrasound has also been incorporated into three separate multiclass machine learning algorithms to predict the degree of coronary disease seen on angiogram showing on average 105.23% improvement in AUC and 27.75% improvement on classification accuracy compared with FRS, SCORE, and ASCVD tools.⁵¹ One group has incorporated novel high-fidelity audio signals from the heart, known as phonocardiogram, with more traditional ECG echocardiogram, Holter monitor, and lab markers in a multimodal approach to predict obstructive coronary disease on angiography (>70% coronary vessel occlusion) with high accuracy (96.67%) in patients presenting with chest pain.⁵² In fact, their model improved in a stepwise fashion, moving from single to multimodal, incorporating all 5 data streams. Although focused outside the coronary space but illustrating the power of using individual data modalities to strengthen the phenotype predictions from other modalities, a cross-modal framework was established using relatively simple ECGs and highly complex cardiac magnetic resonance images to allow imputation of left ventricular structural outcomes using ECG signals.⁵³ We are likely to see expansion of cross-modal integrative models to enhance CAD risk prediction. Sometimes important signals are derived from data not traditionally gathered within a specific medical subspecialty. That is the case for cardiology in learning from ophthalmology where several studies have illustrated that models built from retinal fundoscopic imaging, alone or in concert with clinical data, provide enriched cardiovascular disease risk and event prediction.^54–57 Indeed, the first published foundation model in medicine used self-supervised learning of 1.6 million retinal images to accurately predict the risk of not only multiple eye conditions but also multiple systemic diseases including heart attack, heart failure, stroke, and Parkinson’s disease.⁵⁸ As individuals, we seek to understand the specific factors in our lives that contribute to an elevated risk of disease. This knowledge enables us to focus our efforts on targeted improvements in those areas. While many AI algorithms inform us about the risk, they often operate as “black boxes,” providing little insight into the features influencing their determinations. As advanced multimodal AI models integrate increasingly intricate data for assessing disease risk, it becomes crucial to enhance the explainability of these models.^59,60 This emphasis on transparency is essential for fostering trust and promoting a better understanding of these models on a personal level.

BLOOD PRESSURE MONITORING AND MANAGEMENT

The impact of blood pressure control on future adverse events in cardiometabolic disease including heart attack, stroke, kidney disease, and death cannot be understated. Hypertension remains the most common modifiable risk factor. In fact, in the United States alone, nearly 50% of adults meet criteria for hypertension. Sadly, despite increasing focus on blood pressure education and awareness, in addition to improved accessibility of blood pressure monitors in non-clinical settings an estimated 38% of adults with hypertension are unaware of their disease.⁶¹ Because blood pressure is influenced by a complex interplay of biologic and environmental factors, it is not surprising that roughly 1 in 4 individuals with hypertension are adequately controlled to target goals. The variability in individual responses to common blood pressure medications has been shown to be quite large, illustrating the need for better metrics and predictive aids to map out the best pharmacologic options.⁶² Several groups leveraged AI to determine treatment goals, standard versys intensive blood pressure control,⁶³ and to predict treatment response to specific classes of medications.^64,65 One model predicted up to a 70% reduction in systolic blood pressure from standard of care, though such impressive reductions await prospective validation.⁶⁵ Multimodal AI approaches incorporating the top 50 features of biochemical, dietary, genomic, and metabolic inputs have further expanded our understanding of the biologic underpinnings leading to high blood pressure, opening up pathways to novel pharmacologic and personalized treatment pathways.⁶⁶ Additionally, successful models predicting an individual’s blood pressure status were developed from non-cardiac data streams including retinal fundus images and body composition as measured by bioelectrical impedance analysis.^56,67

In addition to the keen interest in using AI in blood pressure prediction, treatment response, and impact on outcomes, more recently there has been emphasis on understanding real-time blood pressure measurements.⁶⁸ The current gold standard for blood pressure evaluation remains the 24-h ambulatory blood pressure monitor. These are based on oscillometric methods and typically record blood pressure every hour while being worn. The importance of ambulatory blood pressure measurement in comparison to one-off or infrequent clinic blood pressure measurements is well documented.⁶⁹ This is especially true of night-time measurements where those measurements outperformed clinical blood pressure readings by six times in relation to risk of death. However, traditional 24-h ambulatory blood pressure monitoring can be costly and obtrusive. With the technological advances in app development, sensor design, and interconnectivity, there has been significant effort in evaluating the effects of these interventions in randomized controlled trials.⁷⁰ To date, these efforts are promising, though with heterogeneous results. However, the era of wearable, cuffless, and continuous blood pressure monitoring is growing increasingly near with improved accuracy of models incorporating multiple signals derived from photoplethysmography and ECG (pulse arrival time and pulse wave morphology)^71,72 and recently reported piezoelectric sensor arrays.⁷³ These combined bioengineering and machine learning efforts have been reviewed in detail⁷⁴ and highlight the benefits of data-rich AI models. Though, as we have seen with other modalities in the ability to extend their diagnostic capabilities beyond traditional abilities (such as the ECG for predicting left ventricular function), machine learning analysis of sensor-based, continuous blood pressure monitors can illustrate key hemodynamic aspects of the heart (stroke volume, systemic vascular resistance, and cardiac output) that were traditionally captured using echocardiographic or invasive methods.⁷⁵

SLEEP OPTIMIZATION

That poor sleep is associated with poor health outcomes is undisputed.^76,77 Factors identified in the complex relationship between cardiovascular health and sleep include duration, quality (or restfulness), timing, efficiency, and disease patterns of obstructive sleep apnea (OSA). The importance of sleep has been highlighted by the American Heart Association officially recognizing sleep as an essential modifiable risk factor for heart disease in their most recent preventive guidelines.²⁷ While a risk factor itself, sleep factors also contribute to separate independent risks such as diabetes and hypertension.⁷⁸ Concerningly, studies have shown that only 65.2% of adults meet the recommended 7 or more hours of nightly sleep.⁷⁹

Alone, OSA nearly doubles an individual’s risk of cerebrovascular disease.⁸⁰ With an estimated 1 billion worldwide thought to have OSA, vast efforts have been made to leverage AI to enhance OSA diagnosis and better phenotype disease complexity.⁸¹ In fact, a novel, multimodal machine learning approach incorporating data-rich streams from electroencephalogram, ECG, and electromyogram has improved the diagnostic sensitivity for eight individual sleep disorders including insomnia, narcolepsy, sleep disordered breathing, and periodic limb movement.⁸² To provide a more robust risk assessment for 10-year outcomes beyond the pooled cohort equation, which does not account for sleep disturbance or OSA, AI was applied to data from the Sleep Heart Health study.⁸³ This showed improvement over the Framingham risk score with, not surprisingly, the severity of OSA being a strong contributor to risk in their model. Efforts have been made to simplify cardiovascular risk models incorporating OSA variables to the fewest features as possible, now ranging from 9 to 14, without reducing predictive power, though a standard model remains to be established.⁸⁴

Despite the availability of more widespread screening and treatment strategies, underdiagnosis remains a significant problem. The hope is that improved sensor capabilities allowing for robust and sensitive sleep disorder diagnosis at home will reduce the burden on patients for overnight sleep studies. Multiple sensor platforms are currently available for home sleep monitoring in the form of watches, bands, rings, sleeves, and patches.^85,86 Advancements have led to the development of a form-fitting patch that detects not only specific sleep patterns, and via machine learning accurately detects sleep apnea, but also multiple cardiovascular hemodynamic and flow parameters from multimodal inputs.⁸⁷

STRESS AND DEPRESSED MOOD

The bidirectional relationship of psychosocial stress and depression with cardiometabolic disease is well established and reviewed in detail elsewhere.⁸⁸ Appreciating the diversity of populations studied, variability of metrics used to document depressed mood, and the confounding interplay between mental health and additional cardiovascular risk factors, such as smoking, diabetes, obesity, and sedentary lifestyle, the specific degree of risk has been difficult to quantify.⁸⁹ There are emerging genetic data highlighting the shared genetic architecture among diabetes, coronary disease, and depression.⁹⁰ Both American and European medical professional organizations highlighted that depression is highly prevalent in patients with established heart disease and recommended treating it as a risk factor for subsequent events.^91,92 To improve on diagnosis and quantification of stress and depression, the use of AI here has gained significant traction.^93,94 These efforts are supported by the digital health sensors and wearables as well as multimodal models incorporating questionnaire-level data, facial recognition, and audio signals.^95,96 Others have illustrated the improved predictive performance of machine learning models that integrated clinical as well as socio-demographic/lifestyle datasets to model the impact of anxiety and depression to adverse cardiovascular events in patients with type 2 diabetes.⁹⁷ As the technologic infrastructure expands to incorporate real-world data from wearables and sensors with EHR and genetic data, we expect to see improved pathways for prediction, diagnosis, and management for mental health be supported by multimodal AI.

PHYSICAL ACTIVITY AND PERSONALIZED EXERCISE COACHING

Maintaining moderate levels of physical activity throughout the life’s continuum is a cornerstone of cardiometabolic disease prevention.⁹⁸ The World Health Organization recommends that adults achieve at least 150 to 300 min of moderate-intensity or 75 to 150 min of vigorous-intensity physical activity in addition to 2 or more days of resistance training each week.⁹⁹ The benefit of physical activity over sedentary lifestyle as measured by activity trackers has been well established in multiple populations.^100–103 In the All Of Us research program, the combination of EHRs with consumer wearable activity data illustrated a strong inverse relationship for steps per day and obesity, diabetes, hypertension, sleep apnea, and depression in a diverse population of 6,042 participants.¹⁰⁴ Assessment of participants in the UK Biobank who classified themselves as non-exercisers showed that even multiple, brief 1-2-min bursts of activity were associated with reduced cancer and cardiovascular disease mortality.¹⁰⁰

With the continued expansion of wearable monitors in the cardiovascular space, the tracking of multiple and varied parameters related to physical activity has been possible.^105,106 Consumer wearables tracking physical activity and movement as measured by step count were used extensively in studies of cardiovascular disease. Though these technologies continue to mature, to date, observational, interventional, and validations studies in the cardiovascular space using accelerometry were reported with great heterogeneity, lacking consistency regarding reporting and processing methods.¹⁰⁷ These metrics range from vital sign measurements including heart rate, blood pressure, peripheral oxygen saturation, and step count to physiologic signals incorporating glucose, pH, and lactate.¹⁰⁸ Advanced microfabrication techniques are allowing for an increasingly complex array of sensors in a soft, stretchable format, further augmenting the bright future of multimodal AI in exercise physiology.¹⁰⁹ Machine-learning-based methods for classifying specific activities such as sleep, walking, riding a bicycle, or standing and other activities of daily living (ADLs) from patient-level accelerometer data allowed more detailed assessment of the health benefit from all forms of physical activity in participants from a range of ages.^110–113 Some models were able to discriminate (AUC 0.86) patients with and without type 2 diabetes mellitus from accelerometer data.¹¹⁴

The recognition and quantification of various physical activities based on wearable monitoring has been enhanced using machine learning using multiple algorithms of increasing complexity.¹¹⁵ However, machine learning approaches still lag and were only used by a minority of studies (8 of 20) completed from 2018 through 2020 reported in a recent systematic review.¹¹⁵ We expect this trend to reverse, with the majority of future studies likely to rely on AI to some degree. As more complex sensors are used, the use of advanced multimodal algorithms will also increase. Already, using multimodal algorithms has improved the prediction of the intensity of physical activity with the integration of traditional accelerometer data with electrodermal activity, skin temperature, and heart rate.¹¹⁶

Typically, we observe that even well-designed, high-touch lifestyle interventions promoting exercise and healthy behaviors, while successful in the short-term, often lack long-term adherence.^117,118 Accordingly, AI has been adapted to predict adherence to prescribed activity programs with high sensitivity and specificity (AUC 0.90), which can be used to better target individuals at high lapse risk.¹¹⁹ Another approach to address this emphasis has been digital interventions in the form of personalized coaching. These interventions included basic variations in text messaging systems,^120,121 while others incorporated reinforcement learning algorithms to improve the impact of specific messages.^122,123 Other models combined deep learning and semantic ontologies to incorporate personal characteristics, goals, preferences, and individual constraints in physical activity coaching.¹²⁴ AI-assisted chatbots were effective in increasing physical activity and health behaviors, especially when combined with multicomponent interventions.^125,126 Delivery is important as well. When text message versus voice-assisted AI-based coaching methods were evaluated to promote physical activity in cancer survivors, the voice-assisted coaching via a smart speaker proved most effective.¹²⁷ Recent developments focused on the ability of an AI-assisted digital health coach to adapt in real time to increase the personalization of successful health guidance.¹²⁸ There are several studies currently underway leveraging adaptive messaging and activity tracking to promote physical activity and lifestyle interventions with AI in patients with prevalent cardiovascular disease, prediabetes, and type 2 diabetes mellitus.^129–131

With the use of large-scale, deeply phenotyped population data and machine learning approaches, the concept of the digital twin for the personalization of physical activity and coaching for disease prevention, rehabilitation, and competitive athletics has been explored in detail.¹³² While many existing programs exhibit autonomy and interactivity, there is a lack multimodal feedback such as through the integration of haptics and attention to data privacy and security. The desire is strong for effective AI tools to better guide and target patients through specific rehabilitation and preventive programs.¹³³

AI has been shown to improve diagnostic performance, even when restricted to more traditional clinic-based assessments of risk in patients with cardiovascular disease. For exercise treadmill stress tests in predicting obstructive CAD, AI models outperformed conventional metrics with reduced false-positive rate.¹³⁴ Similarly, a machine learning algorithm improved prediction of adverse events in patients undergoing cardiopulmonary exercise testing (CPET).¹³⁵ Moreover, machine learning models that use vector regression of wearable sensor data can predict cardiovascular fitness effectively, potentially rendering traditional CPET unnecessary.¹³⁶

A more robust understanding of the genetic contributions to physical activity and sedentary behavior opens up new layers that may one day inform individualized exercise and activity plans.^137,138 AI has already been used to model relationships between genetic risk markers for cardiometabolic disease and high-resolution phenotypes obtained from wearable sensor and clinical data, showing that blood pressure tracks with markers of activity and obesity and dyslipidemia with sedentary behaviors.¹³⁹ Likewise, specific laboratory test results (especially those related to complete blood count) are predicted by wearable vital signs using AI models.¹⁴⁰ The integration of multimodal data from EHR, genomics, and mobile sensors for physical activity combined with individualized exercise preferences, behavior responses to specific nudges, and e-coaching is expected to improve uptake and adherence to lifestyle interventions for improved physical activity.

INNOVATIONS IN DIABETES AND GLUCOSE CONTROL

Type 2 diabetes has long been established as significant risk factor for ASCVD and is one of the key areas of emphasis for disease prevention.²⁶ While expanded treatment modalities, integrated care, and improved education have resulted in a decreased incidence of adverse cardiovascular outcomes associated with diabetes over time, the risk remains significantly elevated over non-diabetics.¹⁴¹ The formal diagnosis of type 2 diabetes mellitus is traditionally made when an individual’s fasting glucose levels or HbA1c reaches prespecified thresholds, specifically when fasting glucose is more than 126 mg/dL or HbA1c is found to be at or above 6.5%.¹⁴² However, the pathophysiologic impact of elevated blood glucose does not necessarily wait to start at those given values, like an on/off switch, and we have learned that metabolic dysregulation leading to higher than desired post-prandial glucose spikes and daily glucose variability are red flags as well.¹⁴³ Remarkably, an estimated 14.7% of the US adult population has blood glucose levels high enough for the diagnosis of diabetes, and an additional 38.0% have prediabetes.¹⁴⁴ As observed for high blood pressure, nearly 20% of people with diabetes and 80% of people with prediabetes are unaware they have it.

Although public awareness outreach campaigns and community screening efforts have improved, underdiagnosis remains a key issue, and this has not gone unnoticed by the AI field. Recently, medical imaging in the form of CT scans and chest X-rays (CXR) were found to hold clues to diabetes diagnosis. For patients having non-contrast abdominal CTs completed as part of routine colorectal cancer screening, a deep learning model identified patients with type 2 diabetes (AUC 0.85) or at risk of type 2 diabetes (AUC 0.81) using features including low pancreatic attenuation, visceral fat, and intrahepatic fat.¹⁴⁵ Given the high worldwide utilization of CXR over CT scans, another group trained a deep learning model on 271,065 CXRs in patients with linked electronic health data to diagnose type 2 diabetes with an AUC of 0.84 in a testing set of 9,943 CXRs.¹⁴⁶ Validation of this algorithm at an independent health center showed AUC of 0.77. Adding a diabetes prediction score to the reports of these type of radiology studies done in urgent care and emergency room setting would be a low-cost and fully automated effort to improve diabetes screening for many individuals who may have otherwise not been engaged with the healthcare system. For patients already connected to healthcare systems, prediction of patients at high risk for developing diabetes (AUC 0.80) is possible using models that are developed using an array of administrative health data from 2.1 million individuals including demographics, laboratory values, drug benefits, clinic and acute care visits, social determinants of health, and medical records.¹⁴⁷

The genetic contributions to diabetes risk should not be overlooked. PRSs, incorporating the additive contributions of thousands to several million loci across the genome, effectively stratify individuals at low to high risk of diabetes beyond traditional and commonly used markers.^36,148,149 As larger and more ethnically diverse populations are studied, these PRSs improved performance across ancestries.¹⁵⁰ Modeling PRS to diabetes risk was improved by machine learning approaches when used alone, in combination with traditional risk metrics, and when incorporating multiple PRSs, which further enhances the ability to target individuals at the highest risk of diabetes with more effective prevention and treatment strategies.^151,152 Beyond PRS, efforts to incorporate more complex, multimodal inputs from metabolome, proteome, and microbiome have also been successful, identifying individuals at an elevated risk of diabetes who would have otherwise flown under the radar and who benefit the most from early intervention.^153–155

For individuals already diagnosed with diabetes mellitus, AI is being guided to improve treatment strategies and reduce complications.¹⁵⁶ These include impressive multimodal, multi-omic models to understand what pharmacologic therapies best fit for an individual with a specific genetic or microbiome profile and highlight drug-drug interactions.¹⁵⁷ Additionally, improved detection of diabetic retinopathy and acute kidney injury, two common complications of diabetes, was empowered by AI.^57,158 To date, while interventions harnessing the interconnected and wireless capabilities of digital health technologies have shown variable success, though in limited and heterogeneous populations, there is a great opportunity for advanced AI integration.^159,160 As these capabilities for medical AI in this space improve, it is important to note the findings from a survey of 8,420 participants in Denmark; the survey found that while patients were ready to embrace AI technology to augment their disease care, they continued to desire and value human interaction, and that those respondents with diabetes were less open to human replacement.¹⁶¹

A NEW ERA OF PRECISION NUTRITION

Globally, poor diet and suboptimal nutrition were found to contribute to 11 million deaths and significant morbidity, accounting for 255 million disability-adjusted life years, in 2017 alone.¹⁶² An examination of the data taken from the National Health and Nutrition Examination Survey (NHANES) in the US showed that obesity and subsequent diabetes in young adults (aged 20–44 years) increased progressively in the years studied from 2009 to 2020 with over 40% of this population with obesity.¹⁶³ Concerningly, the full public health implications of this disease burden are yet to be understood. Adiposity-based chronic disease is a significant predictor of coronary heart disease, cerebrovascular disease, and heart failure compared with persons of healthy weight, even in the absence of diabetes or prediabetes, calling into the question the concept of metabolically healthy obese.¹⁶⁴ The pathophysiology of weight management remains poorly understood, and the “calories in minus calories out” adage is no longer an acceptable notion, especially considering newly recognized genetic underpinnings of disease. For one obesity PRS comprising 2.1 million SNPs, the highest decile of obesity PRS was associated with progressively increasing risks of higher BMI, odds ratio of 4.2, 6.63, and 14.42 for BMI >40, >50, and >60, respectively.¹⁶⁵ Appreciating the range of inherent genetic risk for unhealthy weight combined with the findings from an impressive meta-GWAS of glucoregulatory genes in 476,326 multiethnic individuals illustrating novel mechanisms of the gut to regulate blood glucose levels specific to individualized genetic markers leaves the door wide open for new discoveries in the field of AI-empowered precision nutrition.^166–168

With greater implementation of continuous glucose monitoring (CGM) technologies, it became evident that a person’s individual blood glucose responses to specific foods was not only highly variable but also quite predictable with the integration of multimodal AI.¹⁶⁹ Although these models were built on data layers from CGM, food logs, anthropomorphic data, physical activity monitoring, and characterization of gut microbiome, other models were built using the microbiome alone to predict the glucose response to various types of bread.¹⁷⁰ The integrative models validated in individuals with and without diabetes show utility for nutritional planning in prevention and management of disease.^171,172 While there are several ongoing studies in the precision nutrition space, including the $156 million NIH-sponsored Nutrition for Precision Health effort targeting 10,000 participants of the All of Us research program, the PREDICT-1 study completed in 1,002 adults in the United Kingdom developed a successful machine learning model to predict both triglyceride and glucose responses to food intake.^173,174 This study showed the importance of nutrient content, microbiome, and environmental (meal timing, sleep, and physical activity) factors on post-prandial glucose responses while a strictly heritable contribution alone from genetics was less predictive than expected. Short (90 day) and medium (6-month) term follow-up studies have shown that various AI-empowered precision nutrition programs successfully lead to reduced HbA1c levels, and in some cases diabetes remission, but they have variable effect on weight loss.^175–178 The technological advances in digital health including CGM that enabled these studies are expanding with newer sensor platforms detecting glucose from sweat rather than the interstitial space.^179,180 Furthermore, AI models of CGM data predicting the effect of other environmental factors such as day of week, holidays, and daily temperature on glucose levels will one day need to be incorporated into individualized guidance strategies.¹⁸¹

The significant impact of the gut microbiome on post-prandial glucose response and its role in the progression of cardiometabolic diseases has underscored the importance of considering the gut microbiome as a potential marker of disease risk. Additionally, this emphasizes the potential for dietary interventions to modify the gut microbiome toward a profile that promotes overall health.^182,183 AI models have shown specifically that insulin resistance is associated with microbiome species variants and that gut microbiome could potentially be used to select for more successful weight loss diets.^184,185 The efforts to better define personal phenotypes based on deeper data did not stop with the gut microbiome but rapidly expanded into a multi-omics assessment of health and disease states.^186,187 Using AI models that integrated variations of metabolomic and/or microbiome profiles, several groups have revealed specific signatures for cardiometabolic disease, the severity of CAD, and acute heart attacks, independent of traditional risk factors.^188–190 Comprehensive multiomic profiling predicted adiposity and modeled likely body weight changes given a specific lifestyle intervention.¹⁹¹ Other multimodal studies used support vector machine modeling of 50 features from gut metabolomics and 76 from brain magnetic resonance imaging to discriminate obese from very obese patients with 90.49% accuracy.¹⁹²

The next phases in precision nutrition involve the integration of personalized dietary strategies in a variety of patient groups and extended-term outcome evaluations. Building upon the proven effectiveness of GLP-1 agonists in achieving substantial weight loss, enhancing diabetes management, and decreasing adverse cardiovascular incidents, it is probable that AI-supported multiomic analysis will aid in identifying the specific patients who would gain from this pharmaceutical approach.^{166,193–196} Additionally, it could potentially offer guidance on a nutritional plan or identify less cost-prohibitive approaches to sustain the enhanced metabolic outcomes while avoiding weekly injections and reducing the need for lifelong therapy.

PULLING IT ALL TOGETHER

Although the impact of cardiovascular disease on morbidity and mortality has reduced on a population level over the past several decades, there remain significant opportunities for improvement. We have seen tremendous and accelerated growth in medical AI, though primarily with single-modality models. However, the introduction of transformers and multimodal techniques unlocked the potential to integrate intricate data sources including medical imaging, labs, clinic notes, genomics, microbiome, and digital sensors. These innovations enable us to attain a more complex and nuanced understanding of an individual’s unique disease progression. As we continue to map out the complex metabolic relationships between diet, physical activity, stress in our environment, personal genetic risks, and medications, we remain consistently underinformed and unable to align personal health goals with the appropriate, long-lasting interventions. It is here that we expect to see multimodal medical AI play a major role, though we still lack the strength and specificity of models for prognosis and prediction of cardiometabolic disease where prospective validation has proved that they are ready for widespread, independent adoption. Certainly, work in the oncology space has been promising, especially in screening for pancreatic ductal adenocarcinoma by CT scan, which highlights the promise for similarly powered AI models in cardiovascular disease.¹⁹⁷

Of course, the advances within medical AI do not come without their share of challenges. Of primary concern is the data quality and quantity. Given the scale and heterogeneity of multimodal healthcare data, improved methods to harmonize, clean, and impute missing data will be needed.¹⁹⁸ Additionally, most of the existing healthcare data are biased, reflecting limited ethnic diversity and unbalanced socioeconomics, and complicated by access to resources and technology.^199,200 Efforts to understand and avoid algorithmic bias resulting from the data flaws are ongoing to avoid further exacerbation of healthcare disparities in the AI era.^201,202 Data privacy and security are paramount, and given the scale of data and need for collaboration between public and private sectors, there will need to be transparency in what types of data are being shared and increased efforts made to avoid data leaks and hacks. We are encouraged by the recent announcements in the US and abroad for a more active role by the government and scientific oversight commissions to establish guardrails for risk management and transparency as AI is adopted more broadly.^203–205 Prospective clinical validation in large, multiethnic populations will be required to build trust in models for both clinicians and patients.^206–208 It is expected that AI will also empower the next generation of digital clinical trials to allow for large-scale, patient-centric assessment of clinical utility.²⁰⁹ Importantly, cardiometabolic risk factor optimization is a long game that ideally starts early in life and is maintained over time. Even despite knowing and understanding best practices and interventions, sustaining a nutritious diet, optimal physical fitness, and healthy weight and avoiding stress, drugs, and maladaptive behaviors is a monumental task. Here, too, we look to AI to assist in mapping out individualized motivational tactics adaptive to changing priorities and phases in life using personal chatbots and virtual advisors. The use of generative AI using large language models has rapidly accelerated and has moved from straight text now to incorporating images and voice. These tools improve productivity, integrate medical information, and even answer patient questions with incredible accuracy but with improved empathy compared with busy care providers.^210–213 Although tools such as GPT-4 are exciting, they are not perfect, and it remains unclear how much trust we can place on the output, especially when it comes to health guidance. This too will continue to improve over time. The next phase in medial AI is incredibly exciting and will no doubt challenge the existing dogma of medical thinking as well as upend a business model that is not only unsustainable but also leaving patients without the support and guidance they need and deserve. This is especially true when it comes to lifelong prevention, which is vital in cardiometabolic disease.