Secondary Prevention of AFAIS: Deploying Traditional Regression, Machine Learning, and Deep Learning Models to Validate and Update CHA2DS2-VASc for 90-Day Recurrence

Jenny Simon, Łukasz Kraiński, Michał Karliński, Maciej Niewada

Backgrounds/Objectives: Atrial fibrillation (AF) confers a fivefold greater risk of acute ischaemic stroke (AIS) relative to normal sinus rhythm. Among patients with AF-related AIS (AFAIS), recurrence is common: AFAIS rate is sixfold higher in secondary versus primary prevention patients. Guidelines recommend oral anticoagulation for primary and secondary prevention on the basis of CHA2DS2-VASc. However, guideline adherence is poor for secondary prevention. This is, in part, because the predictive value of CHA2DS2-VASc has not been ascertained with respect to recurrence: patients with and without previous stroke were not routinely differentiated in validation studies. We put forth a protocol to (1) validate, and (2) update CHA2DS2-VASc for secondary prevention, aiming to deliver a CPR that better captures 90-day recurrence risk for a given AFAIS patient. Overwhelmingly poor quality of reporting has been deplored among published clinical prediction rules (CPRs). Combined with the fact that machine learning (ML) and deep learning (DL) methods are rife with challenges, registered protocols are needed to make the CPR literature more validation-oriented, transparent, and systematic. This protocol aims to lead by example for prior planning of primary and secondary analyses to obtain incremental predictive value for existing CPRs. Methods: The Virtual International Stroke Trials Archive (VISTA), which has compiled data from 38 randomised controlled trials (RCTs) in AIS, was screened for patients that (1) had an AF diagnosis, and (2) were treated with vitamin K antagonists (VKAs) or without any antithrombotic medication. This yielded 2763 AFAIS patients. Patients without an AF diagnosis were also retained under the condition that they were treated with VKAs or without any antithrombotic medication, which yielded 7809 non-AF AIS patients. We will validate CHA2DS2-VASc for 90-day recurrence and secondary outcomes (7-day recurrence, 7- and 90-day haemorrhagic transformation, 90-day decline in functional status, and 90-day all-cause mortality) by examining discrimination, calibration, and clinical utility. To update CHA2DS2-VASc, logistic regression (LR), extreme gradient boosting (XGBoost), and multilayer perceptron (MLP) models will be trained using nested cross-validation. The MLP model will employ transfer learning to leverage information from the non-AF AIS patient cohort. Results: Models will be assessed on a hold-out test set (25%) using area under the receiver operating characteristic curve (AUC), calibration curves, and F1 score. Shapley additive explanations (SHAP) will be used to interpret the models and construct the updated CPRs. Conclusions: The CPRs will be compared by means of discrimination, calibration, and clinical utility. In so doing, the CPRs will be evaluated against each other, CHA2DS2-VASc, and default strategies, with test tradeoff analysis performed to balance ease-of-use with clinical utility.