DARE-ISC model for prediction of 1-year ischaemic stroke risk in the general population and atrial fibrillation patients: a Danish nationwide cohort study

Abstract

Objectives: To develop a risk assessment model (DAnish REgister Ischaemic Stroke Classifier, DARE-ISC) for predicting 1-year primary ischaemic stroke/systemic embolism (SE) in the general population. Secondly, to validate the accuracy DARE-ISC in atrial fibrillation (AF) patients where well-established models and risk scores exist.

Design: Retrospective cohort study. DARE-ISC was developed using gradient boosting decision trees with information from 375 covariates including baseline information on relevant diagnoses, demographic characteristics, registered health-services, lifestyle-related covariates, hereditary stroke components, drug prescriptions and stress proxies.

Setting: Danish nationwide registries.

Participants: All Danish individuals aged ≥18 from 2010 to 2017 (n=35 519 348 person-years). The model was trained on the 2010-2016 cohorts with validation in the 2017 cohort.

Primary and secondary outcome measures: Model optimisation and validation were performed through comparison of the area under the receiver operating characteristic curve (AUC) and average precision scores. Additionally, the relative importance of the model covariates was derived using SHAP values.

Results: DARE-ISC had an AUC (95% CI) of 0.874 (0.871 to 0.876) in the general population. In AF patients, DARE-ISC was superior to the GARFIELD-AF risk model and CHA₂DS₂-VASc score with AUC of 0.779 (95% CI 0.75 to 0.806), 0.704 (95% CI 0.674 to 0.732) and 0.681 (95% CI 0.652 to 0.709), respectively. Furthermore, in AF patients, DARE-ISC had an average threefold and fourfold higher ratio of correctly identified strokes compared with the GARFIELD-AF risk model and CHA₂DS₂-VASc score, as indicated by average precision scores of 0.119, 0.041 and 0.034, respectively.

Conclusions: DARE-ISC had a very high stroke prediction accuracy in the general population and was superior to the GARFIELD-AF risk model and CHA₂DS₂-VASc score for predicting ischaemic stroke/SE in AF patients.

Keywords: cardiac epidemiology; epidemiology; stroke.

Figure 1

(A) The study population and the separation of the cohorts into training, validation and external datasets; (B) the model training phase and how it interrelates with the datasets in figure 1A.

Figure 2

Covariate importance scores for the 30 most important covariates in the 2017 population-based on SHAP values. A high positive SHAP value indicates a large impact on the predicted risk of a stroke while negative values are associated with low risks of stroke. Continues variables are encoloured on the blue-red diverging palette shown in the legend, while categorical values are coloured grey. T1–T10 indicate the lagged time period(s) in years over which the covariates are computed. DISCO, the Danish version of the International Standard Classification of Occupations.

Figure 3

ROC curve and precision-recall curve for DARE-ISC in the full population. Each point on the precision-recall curve is a cut-off point. The axes show the incidence rate of strokes above the cut-off (precision) and the true positive rate (recall) at those cut-offs. The recall is naturally a function of the cut-off, for example, at a cut-off with a 20% precision, the recall is 7.5%; The naive model, random guessing, is equal to the ischaemic stroke/SE incidence rate at all cut-offs; AP is calculated as the weighted mean of precisions across all cut-offs. AP, average precision; DARE-ISC, DAnish REgister Ischaemic Stroke Classifier; ROC, receiver operating characteristic.

Figure 4

ROC curve and precision-recall curve for DARE-ISC in the atrial fibrillation cohort for DARE-ISC, GARFIELD-AF and CHA2DS2VASc. AF, atrial fibrillation; AP, average precision; DARE-ISC, DAnish REgister Ischaemic Stroke Classifier; ROC, receiver operating characteristic; CHA₂DS₂-VASc and GARFIELD-AF are existing stroke prediction models developed for AF patients.