Development of machine learning-based models to predict 10-year risk of cardiovascular disease: a prospective cohort study

Abstract

BackgroundPrevious prediction algorithms for cardiovascular diseases (CVD) were established using risk factors retrieved largely based on empirical clinical knowledge. This study sought to identify predictors among a comprehensive variable space, and then employ machine learning (ML) algorithms to develop a novel CVD risk prediction model.MethodsFrom a longitudinal population-based cohort of UK Biobank, this study included 473 611 CVD-free participants aged between 37 and 73 years old. We implemented an ML-based data-driven pipeline to identify predictors from 645 candidate variables covering a comprehensive range of health-related factors and assessed multiple ML classifiers to establish a risk prediction model on 10-year incident CVD. The model was validated through a leave-one-center-out cross-validation.ResultsDuring a median follow-up of 12.2 years, 31 466 participants developed CVD within 10 years after baseline visits. A novel UK Biobank CVD risk prediction (UKCRP) model was established that comprised 10 predictors including age, sex, medication of cholesterol and blood pressure, cholesterol ratio (total/high-density lipoprotein), systolic blood pressure, previous angina or heart disease, number of medications taken, cystatin C, chest pain and pack-years of smoking. Our model obtained satisfied discriminative performance with an area under the receiver operating characteristic curve (AUC) of 0.762±0.010 that outperformed multiple existing clinical models, and it was well-calibrated with a Brier Score of 0.057±0.006. Further, the UKCRP can obtain comparable performance for myocardial infarction (AUC 0.774±0.011) and ischaemic stroke (AUC 0.730±0.020), but inferior performance for haemorrhagic stroke (AUC 0.644±0.026).ConclusionML-based classification models can learn expressive representations from potential high-risked CVD participants who may benefit from earlier clinical decisions.

Refining the Cardiovascular Health Score in Patients With Type 2 Diabetes

“…In a UK Biobank cohort with more than 470,000 participants and 10 years of follow-up, the performance of a ML model was compared against certain traditional CV risk scores. 14 The observed area under the receiver operating characteristic curve of the ML model was 0.762, which was higher than the area under the receiver operating characteristic curves of existing clinical CV risk scores ( P < 0.001). In the Silesia Diabetes-Heart Project, we showed used ML algorithms could identify patients with diabetes at a high risk of new CV events based on a small number of interpretable and easy-to-obtain patients’ parameters.…”

Künstliche Intelligenz in der kardiovaskulären Medizin – Status und Perspektiven

“…Bisherige Algorithmen zur Vorhersage von Herz-Kreislauf-Erkrankungen (CVD) wurden anhand von Risikofaktoren erstellt, die weitgehend auf empirischem klinischem Wissen beruhen. In einer aktuellen Studie mit über 473 000 Teilnehmern wurde versucht, Prädiktoren über Algorithmen des maschinellen Lernens (ML) einzusetzen, um neuartige Modelle zur Vorhersage des kardiovaskulären Risikos zu entwickeln [9]. Auf der Basis von 645 klinischen Variablen wurden über Algorithmen des maschinellen Lernens Prä-diktoren identifiziert, die eine valide Risikovorhersage für das 10-Jahres-Ereignis einer CVD modellieren können.…”

Meta-Prediction of Coronary Artery Disease Risk

“…Other attempts have failed to perform feature selection in the context of model performance, again leading to minimal improvement relative to standard clinical scores 51 , 52 . For example, You et al developed an ML model for 10-year incident CAD prediction using 645 candidate variables prioritized on the basis of multicollinearity with one another rather than contribution to predictive performance 53 . None of these prior approaches contemplate meta-prediction, which as we will demonstrate, results in the generation of the most important predictive features, which capture hidden unmodifiable risk status not necessarily expressed in biometrics and diagnoses at baseline, and critically, resulting in the identification of at-risk sub-groups with differing risk reduction benefit from standard clinical interventions.…”