Machine learning in precision diabetes care and cardiovascular risk prediction

Abstract

Artificial intelligence and machine learning are driving a paradigm shift in medicine, promising data-driven, personalized solutions for managing diabetes and the excess cardiovascular risk it poses. In this comprehensive review of machine learning applications in the care of patients with diabetes at increased cardiovascular risk, we offer a broad overview of various data-driven methods and how they may be leveraged in developing predictive models for personalized care. We review existing as well as expected artificial intelligence solutions in the context of diagnosis, prognostication, phenotyping, and treatment of diabetes and its cardiovascular complications. In addition to discussing the key properties of such models that enable their successful application in complex risk prediction, we define challenges that arise from their misuse and the role of methodological standards in overcoming these limitations. We also identify key issues in equity and bias mitigation in healthcare and discuss how the current regulatory framework should ensure the efficacy and safety of medical artificial intelligence products in transforming cardiovascular care and outcomes in diabetes.

Keywords: Artificial intelligence; Cardiovascular disease; Diabetes; Digital health; Machine learning; Personalized medicine; Prediction.

Fig. 1

Overview of commonly used algorithms in medical machine learning

Fig. 2

Discrimination, calibration, and net clinical benefit. The comprehensive evaluation of a predictive model requires the simultaneous evaluation of its discrimination, calibration, and incremental value beyond the current standard-of-care. A The area under the receiver operating characteristic curve (AUROC) reflects the trade-off between sensitivity (true positive rate) and specificity (1-false positive rate) at different thresholds and provides a measure of separability, in other words the ability of the model to distinguish between classes (0.5 = no separation, 1 = perfect separation). B Models with similar AUROC may exhibit different behavior when the prevalence of the label varies. The precision–recall curve demonstrates the trade-off between the positive predictive value (precision) and sensitivity (recall), and illustrates how the area under the curve may vary substantially as the prevalence of the label of interest decreases from 50 to 5%. C Models with similar AUROC may also differ in their calibration. A model with good calibration (i.e. blue line) makes probabilistic predictions that match real world probabilities. On the other hand, the model shown in orange underestimates and overestimates risk at lower and higher prediction thresholds, respectively. D Finally, models should be compared against established standard-of-cares while incorporating clinical consequences and comparing the net clinical benefit across varying risk levels to established or no risk stratification approaches. Curves were generated using synthetic datasets for illustration purposes

Fig. 3

Phenomapping-derived tools for personalized effect estimates. Phenomaps enable a visual and topological representation of the baseline phenotypic variance of a trial population while accounting for many pre-randomization features. As shown in an analysis of the Canagliflozin Cardiovascular Assessment (CANVAS) trial [138], a phenomap representation of all enrolled patients shows that the study arms are randomly distributed in the phenotypic space (A). Through a series of iterative analyses centered around each patient’s unique phenotypic location, a machine learning model can learn phenotypic signatures associated with distinct responses to canagliflozin versus placebo therapy (B, C). An extreme gradient boosting algorithm trained to describe this heterogeneity in treatment effect in CANVAS successfully stratified the independent CANVAS-R population into high- (D) and low-responders (E). Panels reproduced with permission from Oikonomou et al. [12]

Fig. 4

Machine learning for predictive enrichment of randomized control trials. Machine learning can be used to guide adaptive clinical trial design though data-driven inference and predictive enrichment. Traditional fixed trial designs do not allow modifications in the patient population, whereas sample size adaptations only allow interim revisions in the power calculations and target sample sizes based on the accumulating rate of primary outcome and safety events. In trials whereas there happens to be clinically meaningful heterogeneity in the treatment effect, a priori inclusion of machine learning, data-driven inference may provide early signals of heterogeneous benefit or harm and a reference for adaptive predictive enrichment. This approach can optimize the trial’s efficacy, shorten its duration, minimize its costs, maximize inference, and ultimately ensure safety for the study participants. ML machine learning

Fig. 5

Explainability and interpretability of medical machine learning. Broadly speaking, more complex algorithms demonstrate better performance when dealing with complex tasks and data inputs. For instance, the recognition of cardiomyopathy using echocardiographic videos may require a deep learning algorithm to model the full extent of temporal and spatial features that carry diagnostic value, whereas predicting the risk of re-admission using electronic health record data may be modelled using generalized linear models. Simpler models, such as decision trees and linear models are intuitive and interpretable, whereas ensemble and neural network-based methods are too complex for the human mind to fully understand. Explainable artificial intelligence (XAI) methods aim to bridge this interpretability gap by offering direct or indirect insights into the inner workings of complex algorithms