Machine learning in precision diabetes care and cardiovascular risk prediction

Though AI and ML are inextricably linked, they are not identical terms [22, 23]. Artificial intelligence (AI) describes the ability of a machine to perform tasks that are typical of human intelligence, such as understanding natural language, problem-solving, or creative tasks like generating images and text. On the other hand, the process through which an AI system acquires this ability, learning and improving from experience and observed data to make predictions about new or unseen cases, is called machine learning (ML).

The specific step during which a model learns from data is also known as training, whereas the respective dataset is referred to as training set. Here, the model makes predictions and subsequently adjusts its parameters based on a metric that quantifies how good or bad the predictions are (loss function). It is typical that during training, the model will be applied to an unseen group of observations (validation set) to get a more reliable assessment on performance on unseen data. Further testing in external sets drawn from geographically and temporally distinct populations can serve to solidify claims about external model validity [24].

The learning process can be supervised or unsupervised [22]. Supervised learning describes an iterative process that selects relevant input features and then assigns weights to link the input data to a given value (regression) or class (classification). Unsupervised learning, on the other hand, analyzes and clusters unlabeled datasets by identifying similarities and dissimilarities between data points, therefore uncovering hidden patterns in the data. These two approaches should be considered complementary and are often used in conjunction to address distinct problems. Supervised learning can be used to better predict future cardiovascular risk (regression), or the presence of diabetic retinopathy (classification), whereas unsupervised approaches can be used to identify distinct phenotypic clusters of patients with diabetes with differences in baseline risk, prognosis, and treatment response.

Building on these concepts, self-supervised learning (SSL) processes unlabeled data to create key representations that can facilitate downstream tasks [25]. In principle, SSL closely resembles unsupervised learning since it is applied to unlabeled data. However, instead of focusing on tasks like clustering, SSL attempts to solve tasks traditionally addressed through supervised learning, such as classification and regression [26]. We [27, 28], and others [29, 30], have shown that this is a powerful method to train clinical models, especially when there is a paucity of high-quality labels.

Whether supervised or unsupervised, the ML process requires a set of rules and statistical techniques that can learn meaningful patterns from data, known as algorithms. A representative list of ML algorithms used in medical applications is shown in Fig. 1. Ranging from linear regression to deep learning algorithms, these vary substantially in their ability to model complex data, interpretability, and performance [22]. Further, they can be adapted to model not only cross-sectional or short-term outcomes, as done with logistic regression, but also long-term predictions through survival analysis, similar to Cox regression modeling which has been widely used to estimate CVD risk in the US [31], UK [32], and Europe [33]. Some notable examples include (survival) random forests and deep learning algorithms, which have all been adapted to model long-term hazards [34‐36].

The comprehensive evaluation of the performance of a predictive ML model requires an integrated assessment of discrimination, calibration, and clinical benefit (Fig. 2; a detailed table of metrics used in classification and regression tasks, along with their strengths and weaknesses is also presented in Additional file 1: Table S1) [37]. Commonly used discrimination metrics, such as the Area Under the Receiver Operating Characteristic curve (AUROC) and the Area Under the Precision–Recall Curve (AUPRC) describe the model’s ranking of individual predictions and ability to discriminate between different classes. However, AUROC may not offer a complete description of the model’s performance, particularly in imbalanced datasets. For instance, a model with 95% sensitivity and specificity screening for a rare label (e.g., screening for type 1 diabetes mellitus in the community, prevalence ~ 0.55%), a positive prediction is more likely to be false positive than true positive. Calibration assesses the agreement between the predicted and observed outcomes across the entire range of predictions [38]. Two models may have a similar AUROC but may differ substantially in their calibration performance, a crucial difference that may often impact clinical decision-making if predictions consistently overestimate or underestimate risk. Guidelines on good research practices in prediction modeling suggest that for any given model both discrimination and calibration are reported [39]. Moreover, a complex model may have good discriminatory performance but lack incremental value beyond a simpler or established model, something that can be further evaluated by metrics such as the Net Reclassification Improvement (NRI) and Integrated Discrimination Improvement (IDI) [37]. Finally, good discrimination and calibration do not necessarily translate into net clinical benefit for a specific clinical application. In this setting, decision curve analysis (DCA) can be used to assess the net benefit of a model across a range of potential thresholds for action, weighing the benefit versus harm versus alternative risk stratification approaches [40].

Human end-users may feel reluctant to use what they cannot understand. Interpretability and explainability describe two closely linked yet slightly different concepts [41, 42]. Interpretability refers to the extent to which a human can understand how features are combined to make predictions. This is particularly desirable when training a model to understand its general behavior and identify potential sources of bias. Explainability is a property of ML models that describes the extent to which the inner workings of a model can be explained in human terms, that is, understanding why a model made a specific prediction (often on a case-by-case basis). This is important for the end-user and has practical and regulatory implications. A model can be interpretable without being explainable, and vice versa; however, ideally, a model should be both [43].

Both the American Diabetes Association (ADA) [44], and the United States Preventive Services Task Force (USPTF) [45] have emphasized the importance of early screening for pre-diabetes and diabetes among asymptomatic adults to ensure timely diagnosis and prevent downstream diabetes complications and its sequalae. Current guidelines recommend screening for pre-diabetes and type 2 diabetes with an informal assessment of risk factors or a validated risk calculator among all asymptomatic adults and testing among adults of any age who are overweight or obese and have one or more risk factors [44]. Several risk scores have been proposed for a more personalized risk assessment of type 2 diabetes, such as the American Diabetes Association questionnaire, a logistic regression model trained in National Health and Nutrition Examination Survey (NHANES), Atherosclerosis Risk in Communities (ARIC) and Cardiovascular Health Study (CHS) studies with a reported AUROC of 0.79 to 0.82 for undiagnosed diabetes among U.S. adults aged 20 years or older [46], the Australian type 2 diabetes risk Assessment Tool (AUSDRISK) to predict the incidence of type 2 diabetes over 5 years among participants 25 years or older (AUROC of 0.78) [47], and the Cambridge Risk score to detect cross-sectionally elevated HbA1c levels among individuals aged 45 years (AUROC of 0.84 for HbA1c 7% or greater) [48]. The moderate accuracy of these tools in addition to concerns about their external validity [49] has prompted researchers to explore whether targeted screening could be improved through ML of structured and unstructured data [5‐11].

In NHANES, an XGBoost classifier based on 123 variables showed an AUROC of 0.86 in detecting the presence of an established diabetes diagnosis, though the performance dropped to 0.73 when detecting undiagnosed diabetes adjudicated based on abnormal laboratory findings [50]. In a retrospective analysis of 16 predictors from routine health check-up data of 277,651 participants from Japan, a light gradient boosting machine algorithm was able to predict the 3-year incidence of diabetes with an AUROC of 0.84, demonstrating significantly improved performance compared with a logistic regression model for large training populations of 10,000 patients or more [6]. Another ML algorithm built using EHR and administrative healthcare data from Canada reportedly identified type 1 diabetes cases with 87.2% sensitivity and 99.9% specificity [51].

Moving beyond EHR models relying on administrative datasets, an analysis of 1262 individuals from India showed that an XGBoost algorithm using ECG inputs had excellent performance (97.1% precision, 96.2% recall) and good calibration in detecting type 2 diabetes and pre-diabetes [7]. Furthermore, the integration of a genome-wide polygenic risk score and serum metabolite data with structured clinical parameters using a random forest model resulted in improved type 2 diabetes risk prediction in a Korean cohort of 1425 participants [8]. Several other studies have also defined metabolomic [52] and proteomic signatures for identifying diabetes or insulin resistance [9]. Integrative personal omics profiles (iPOP) that combine genomic, transcriptomic, proteomic, metabolomic, and autoantibody profiles from a single individual over several months can also be harnessed to connect genomic information with dynamic omics activity, describe host-microbiome interactions, and describe personal aging markers, thus risk stratifying various medical risks, including type 2 diabetes [53‐57].

Finally, imaging-based biomarkers are also being studied, with the development of DL models that can automatically segment and output measurements of pancreatic attenuation, volume, fat content, fractal dimension, and parameters associated with visceral adiposity and muscle attenuation/volume. For instance, in a retrospective cohort of 8992 patients undergoing screening CT colonography, a DL model of the above phenotypes could predict the future incidence of diabetes with an AUROC of 0.81 to 0.85 [10].

The traditional classification of diabetes into type 1 and type 2 does not fully capture the complex and highly heterogeneous nature of the condition. From the heterogeneity of the islet microenvironment to the diversity of pathophysiological endotypes that span multiple demographic groups, diabetes mellitus affects a diverse group of patients with distinct molecular underpinnings that require individualized approaches to therapy [58]. Multiomic signatures may provide insights into the interaction of a patient’s genome, phenome, and environment [53‐57], but are often hard to measure at scale.

Unsupervised ML techniques using routinely available EHR computable phenotypes can provide valuable data-driven inference about distinct phenotypic clusters. In longitudinal DL-based clustering of 11,028 patients with type 2 diabetes using a kernelized autoencoder algorithm that mapped 5 years of data, there were seven phenotypic clusters with distinct clinical trajectories and varying prevalence of comorbidities (i.e., hypertension, hypercholesterolemia) or diabetic complications [13]. In a separate analysis of 8,980 patients with newly diagnosed diabetes from Sweden, k-means and hierarchical clustering revealed five replicable clusters with significant differences in the observed risk of diabetic complications [59]. A separate analysis of 175,383 patients with type 2 diabetes further identified 20 frequent comorbidity clusters, and using Bayesian nonparametric models demonstrated a complex and dynamic interrelationship between diabetes-related comorbidities and accelerated disease progression [60].

Despite this, computable definitions can vary significantly in their ability to capture the respective phenotypes. Such differences can have a substantial impact on model performance even within the same center. In an analysis of 173,503 adults from the Duke University Health System, the concordance of variable definitions of diabetes (with or without ICD-9-CM codes) ranged from 86% to as low as 50% [61].

Finally, DL approaches, such as natural language processing, can be used to screen large hospital registries to monitor the quality of care. As shown in an analysis of 33,461 patients with diabetes from 2014 to 2020 in Northern California, a natural language processing approach accurately identified statin nonuse (AUROC of 0.99 [0.98–1.00]) and highlighted patient- (side effects/contraindications), clinician- (guideline‐discordant practice), and system-centered reasons (clinical inertia) for statin nonuse, with notable variation by race and ethnicity [62].

Diabetes is associated with a range of micro- and macro-vascular complications [3]. Given their simplicity and wide availability, fundoscopic images were used in some of the earliest DL models in medicine, predicting diabetic retinopathy with performance matching that of expert readers [11, 63, 64]. This has opened the way for efficient screening of both diabetes, and diabetes-related chronic kidney disease and retinopathy in settings with limited resources, as demonstrated in real-world implementation studies in Thailand and India [65, 66].

While current guidelines endorse routine screening for microvascular complications, there is a paucity of data to support the routine screening for macrovascular complications in asymptomatic individuals [67]. Diabetes has traditionally been regarded as an atherosclerotic CVD (ASCVD) equivalent [44]. Non-invasive cardiovascular imaging approaches, such as measurement of coronary artery calcium (CAC) [68], coronary computed tomography angiography [69], or functional testing [67], are often used to further risk stratify patients with diabetes and diagnosed subclinical CVD. However, such approaches are costly and hard to implement for population-level screening.

ML approaches may support more efficient screening of CVD in this population. In an analysis of NHANES, logistic regression, SVM, XGBoost and random forest models, as well as an ensemble of the four, showed comparable performance in detecting CVD among all-comers with an AUROC of 0.81 to 0.83 [50]. In a separate single-center study from China, training a model to predict the co-occurrence of coronary heart disease and diabetes using 52 structured features in 1273 patients with type 2 diabetes resulted in an AUROC of 0.77–0.80; however, this dropped to 0.7 in an independent dataset, highlighting the challenges in the generalizability of such tools when trained in single-center cohorts [70]. In a retrospective analysis of administrative data from Australia, investigators combined ML techniques with a social network analytic approach to define the disease network for patients with diabetes with or without CVD and identify discriminatory features for CVD presence, with a reported overall AUROC of 0.83 for the random forest classifier, only dropping to 0.81 for a logistic regression model [71]. Overall, it appears that in most cohorts there were minimal gains from using more complex and less interpretable algorithms compared to standard logistic regression. Moreover, many studies do not report the incremental value beyond established risk scores such as the pooled cohort equations, which prevents a reliable assessment of their net clinical value.

Predictive models can further be customized for more specific cardiovascular conditions, such as congestive heart failure, and model survival analyses incorporating time-to-event outcomes. In a post hoc analysis of the ACCORD trial, a non-parametric random survival forest model predicted the risk of incident heart failure, with a C-statistic of 0.77 [72]. A parsimonious five-feature integer-based score based on this model maintained moderate discriminatory performance in both ACCORD (Action to Control Cardiovascular Risk in Diabetes) and an independent test set from Antihypertensive and Lipid-Lowering Treatment to Prevent Heart Attack Trial (ALLHAT) with an AUROC of 0.74 and 0.70, respectively. In a separate retrospective analysis of an EHR-based cohort of patients with diabetes undergoing cardiac testing, a deep neural network survival method resulted in an AUROC of 0.77 for incident heart failure [36].

Many of these studies should be interpreted with caution. First, as shown in a recent systematic review of predictive models for the detection of CVD among patients with diabetes, there is often a high risk of bias in several studies and poor adherence to standardized reporting guidelines [73‐75]. Second, retrospective analyses of single-center cohorts and administrative claims are also prone to perpetuating biases and inequities in healthcare since patients who have better access to healthcare are more likely to utilize such resources and be represented in administrative claims datasets [76].

As the focus shifts from the secondary prevention of diabetes-related complications to earlier prevention in the community, various digital health technologies emerge that can be deployed at scale and minimal cost. Large language models (LLM) have already led to smart conversational agents (“chatbots”), such as ChatGPT, which are freely accessible to most individuals with internet access. Such models are task-agnostic and have been shown to provide “concise”, “well-organized” and easy-to-understand instructions to a series of questions regarding diabetes self-management, albeit with occasional factual inaccuracies [77]. Such tools could be incorporated into existing digital healthcare platforms that combine glucometer, bioelectrical impedance, blood pressure, activity, and AI-derived nutritional analysis data to improve glycemia and weight loss among patients with type 2 diabetes [78].

In the same vein, the rapid uptake in the use of wearable devices and smartwatches has led to AI-enabled solutions to optimize glycemic management. These applications have built on an expanding body of research highlighting the value of AI-ECG (both using 12-lead or 1-lead signals) in detecting subclinical forms of cardiomyopathy and arrhythmias [79‐81]. Recent work from our lab has further expanded these approaches to the use of ECG images [82, 83] and single-lead wearable signals [84], enabling the scaling of such technologies to low-resource settings and to ubiquitous data streams.

In diabetes, AI-ECG-guided monitoring through customized DL models has shown promise in detecting hypoglycemic events [85, 86], and is currently being studied in prospective studies [87, 88]. In one of the pilot studies, investigators trained personalized models using a combination of convolutional (CNN) and recurrent neural networks (RNN) for each participant using data collected over the run-in period, followed by subsequent testing in the same patient. This combination of distinct model architectures takes advantage of the distinct strengths of each model type, with CNN learning hierarchical, abstract representations of the input space, whereas RNN learns sequence patterns across time [85]. Similar concepts have been applied to continuous glucose monitoring (CGM). Here, multi-modal data integration from CGM devices, meal or insulin entries and sensor wristbands has shown promise in detecting hypo- or hyperglycemic events in patients with type 1 diabetes in both simulation [89], as well as small real-world prospective studies [90]. While such technologies have the potential to democratize access to high-value care, we have shown that as of 2020 use patterns suggested disproportionately lower use among individuals with or at risk of CVD than those without CVD risk factors, with fewer than 1 in 4 using such devices [91, 92].

Randomized controlled trials (RCTs) represent the methodological and regulatory gold-standard to test the efficacy and safety of new therapies [93]. However, RCTs traditionally report an average treatment effect (ATE) which does not adequately describe the individualized benefit for each unique patient profiles [94]. The detection of reliable heterogeneous treatment effects (HTE) is limited by the fact that in outcomes trials participants get assigned to one arm (thus the “counterfactual” is never observed). In addition, most trials lack statistical power to detect subgroup differences [95].

One way to explore such differences is through a priori or post hoc-defined clinical subgroups. For instance, it has been shown that sex and body mass index are associated with heterogeneity in the treatment and safety signal of thiazolidinediones and sulfonylureas [96], whereas insulin resistance has been associated with significant differences in the glycemic response to dipeptidyl peptidase 4 (DPP-4) inhibitors [97]. Such subgroups, however, rely on simplistic subgroup definitions, and may not accurately reflect the phenotypic diversity seen in clinical profiles and treatment response. Various statistical approaches have been employed to identify such complex effects, with methods ranging from unsupervised clustering [98] to causal forests and meta-learners (i.e. X-learners), algorithms that can use any supervised learning or regression method to estimate the conditional average treatment effect [94]. In applying this approach in ACCORD and Veteran Affairs Diabetes Trial (VADT), investigators described eight subgroups in which the differences in major adverse cardiovascular events ranged from as low as -5.1% to as high as 3.1% [99].

Most of these approaches focus on the absolute risk reduction of a therapy, which reflects both the relative treatment effect as well as a patient’s baseline risk. We previously developed and tested a phenomapping-based method that creates multidimensional representations of a population based on the full breadth of pre-randomization phenotypes. Over a series of in silico simulations that account for the unique phenotype of each participant relative to all other participants, an ML algorithm learns signatures that are consistently associated with a higher or lower relative treatment effect. In representative applications, our approach has reproduced a beneficial association between the use of anatomical as opposed to functional testing in patients with diabetes and chest pain [100, 101], and highlighted heterogeneity across phenotypes in the cardiovascular benefits of canagliflozin [12] as well as intensive systolic blood pressure control [102] (Fig. 3). However, prospective validation of any post-hoc comparisons is required to inform treatment decisions.

Furthermore, ML can be used to guide the design of adaptive clinical trials, guiding protocol modifications based on accumulating data [103] (Fig. 4). The need for methodological innovation in this space has been embraced by the United States Food and Drug Administration (FDA) [104]. For instance, ML-driven approaches of individualized predictive benefit could be integrated into interim analyses to prioritize randomization of patients with a higher expected net clinical benefit from the studied intervention [105]. In a simulation of real-world clinical trial data from Insulin Resistance in Stroke study (IRIS) [106], and Systolic Blood Pressure Intervention Trial (SPRINT) [107] an ML-informed strategy of adaptive, predictive enrichment enabled a consistent reduction in the number of enrolled participants, while preserving the original trial’s effect [105].

Modern RCTs are both resource and time-intensive [108], particularly when evaluating the effects of novel therapies on major clinical endpoints [109‐111]. Federated analytic approaches that utilize large-scale, multinational, real-world databases, such as the Large-scale evidence generation and evaluation across a network of databases for type 2 diabetes mellitus (LEGEND-T2DM) initiative, are currently underway to enable comparative effectiveness analysis through both traditional and ML-driven big data approaches [112, 113].

In the ML literature, it is widely recognized that a priori knowledge of an optimal algorithm is challenging for any given task. This often depends on the underlying dataset, the performance metric utilized, whereas for any given algorithm performance may still vary with tuning of model-specific hyperparameters (values used to control the training process that are external to the model and cannot be computed from the data). Second, there is a common misconception that for any given dataset, complex models will always outperform less complex ones. This may be true for tasks involving unstructured data, in particular biomedical images and videos, where DL models enable automatic learning and hierarchical combination of key spatial as well as temporal features (rather than relying on hand-engineered features), as well as models that are robust to variations, scalable and transferrable across tasks [114]. However, for structured datasets, such as databases of clinical information used for predictive modeling, the performance of easily interpretable models, such as logistic regression, is often comparable to that of complex extreme gradient boosting or neural network methods [115]. Third, complex models are susceptible to learning noise that may not generalize to a new dataset (overfitting) [116]. In this context, careful consideration of the available data and training plan (i.e., cross-validation), as well as strict separation of training and testing datasets, are warranted to maximize the external validity of new models.

The training of any clinical model should not stop at the time of deployment but rather continue by incorporating real-time data and updating its parameters to prevent a drift in performance when used in a different clinical, geographical, or temporal setting [117]. A notable example here is a widely implemented and proprietary EHR-embedded sepsis prediction model whose external performance was substantially different from the one reported by the model’s creators (AUROC of 0.63 from 0.76–0.83) when deployed across independent hospitals with heterogeneous patient populations [118].

To bridge the interpretability gap of complex “black box” algorithms, various approaches have emerged including Local Interpretable Model-agnostic Explanations (LIME) and SHapley Additive exPlanations (SHAP) [119, 120] (Fig. 5). Though such approaches are often imperfect, ensuring model explainability is not only a regulatory recommendation, but also enhances the adoption of the model in the real world. In a survey of 170 physicians, greater explainability of ML risk calculators was significantly associated with greater physician understanding and trust of the algorithm [121].

Clinical predictions rarely rely on a single factor and are most often multivariable by design. In 2015, to provide a standardized framework for the creation and reporting of such statistical models, the transparent reporting of a multivariable prediction model for individual prognosis or diagnosis (TRIPOD) statement was published [39]. This followed multiple reports, including systematic reviews of risk prediction models in type 2 diabetes, that showed widespread use of poor methods and reporting [122, 123] contributing to “avoidable waste” in research evidence [124]. Unfortunately, several subsequent reviews have shown poor adherence to these standards, particularly among studies using ML algorithms [125]. The increasing adoption of ML algorithms in clinical prediction modeling, despite the lack of clear incremental value beyond simpler methods such as logistic regression in most settings [115, 126], has prompted the original TRIPOD authors to update their statement (TRIPOD-AI, see [127]) researchers, clinicians, systematic reviewers, and policy-makers critically appraise ML-based studies. ML models should generally be used when processing large amounts of multi-dimensional or complex inputs (e.g. time-series from wearables, videos etc.), whereas head-to-head comparisons to traditional statistical models should be provided when feasible to assess the trade-off between performance, complexity, and interpretability.

Since ML models learn from existing data and care patterns, they can perpetuate human and structural biases [128, 129] (Table 1). Careful evaluation of the historical training data for health care disparities, ensuring that historically disadvantaged subgroups have adequate representation, review of model performance across key subgroups, and incorporating feedback from key stakeholders and patient representatives are some approaches that can be taken to mitigate bias [76, 128]. This is not a straightforward task and requires caution when assessing for confounders [130], since ML models have shown the ability to identify features such as race even when blinded to such labels [131].

While clinical decision support (CDS) tools used to assist with electronic patient records, administrative tasks, medical device data, and models aimed at supporting a healthy lifestyle fall outside the FDA’s definition of a “device” [132], most diagnostic and predictive clinical models and CDS are regulated under the “Software as Medicine Device (SaMD)” umbrella and categorized as class I (low-risk), II (moderate-risk), III (high-risk) [20]. In the US, the FDA mandates that class III devices generally undergo a longer and more in-depth review process, known as pre-market authorization. However, for lower or intermediate-risk devices (class I and II), alternative pathways exist [20]. The 510(k) pathway requires manufacturers to show that the risk presented by their device is no greater than that of a substantially equivalent predicate device [20], whereas the de novo pathway is designed for class I or II devices without predicates [133]. While built to accelerate the regulatory process, the latter two pathways have been criticized on certain occasions for facilitating the clearance of devices based on faulty predicates [134]. The regulatory process is different in Europe, where lowest risk devices (class I) are the responsibility of the manufacturer, whereas class II and III devices are processed in a decentralized way through private “Notified Bodies”. Of the 124 AI/ML-based devices approved in the USA and Europe between 2015 and 2020, 80 were first approved in Europe [20]. The European Union’s General Data Protection Regulation (GDPR) further lists explainability as a requisite for any medical AI application [135]. Despite the above, there is no clear consensus as to whether regulatory bodies should require RCT-level evidence to support the effectiveness and safety of their proposed AI tools, even though recent studies have demonstrated the feasibility of testing the net clinical benefit of AI-based ECG and echocardiographic models in the context of pragmatic RCTs [136, 137].