Sep 2026· Int. J. Medical Informatics· Vol 217, pp.
106504
· 0 citations· 63 references
MedicineComputer Science
TL;DR
In light of the pervasive methodological limitations identified, including high analytic risk of bias, absence of external validation, and lack of model interpretability, claims of ML superiority over CHA2DS2-VASc must be interpreted with caution.
Abstract
Background
Atrial fibrillation (AF) is the most common sustained cardiac arrhythmia and confers a four to fivefold increase in ischemic stroke risk, accounting for approximately 15 - 20% of all stroke events globally. Despite this burden, the predominant risk stratification tool, the CHA2DS2-VASc score, achieves only modest discrimination, constrained by its static, additive architecture that cannot capture the nonlinear, high-dimensional interactions inherent in real-world electronic health record (EHR) data. This evidence gap creates a dual clinical hazard: under-anticoagulation in high-risk patients and unnecessary bleeding exposure in those whose risk is overestimated. This study aimed to systematically evaluate the predictive performance, methodological rigor, and clinical readiness of machine learning (ML) models derived from EHR data for the prediction of ischemic stroke in patients with AF.
Methods
A systematic search of PubMed, Embase, Scopus, and Web of Science was conducted from inception through September 2025, following PRISMA 2020 guidelines. Studies were eligible if they developed or validated ML models for ischemic stroke prediction using EHR data in adults with AF and reported at least one quantitative performance metric. Methodological quality was assessed using the PROBAST and TRIPOD-AI frameworks.
Results
Eight studies (2017 to 2024) encompassing 809,523 patients across seven countries were included. Supervised ensemble methods consistently outperformed CHA2DS2-VASc, with AUROCs ranging from 0.66 to 0.91 versus 0.54 to 0.68 for the traditional score. However, performance varied substantially: several models achieved only marginal gains (AUROC 0.63 - 0.69), and the AUROC range reflects pronounced heterogeneity rather than uniform superiority. Critical barriers persist - only one study performed external validation; fewer than half applied explainable AI techniques; class imbalance was rarely addressed; and 88% of studies received a high risk of bias rating in the analysis domain under PROBAST, a finding that substantially limits confidence in the reported performance estimates.
Conclusion
In light of the pervasive methodological limitations identified, including high analytic risk of bias, absence of external validation, and lack of model interpretability, claims of ML superiority over CHA2DS2-VASc must be interpreted with caution. While ML models demonstrate potential discriminative improvements, current evidence is insufficient to support clinical adoption. Translating algorithmic promise into bedside impact requires dynamic longitudinal modeling, rigorous multisite external validation, transparent risk attribution, and prospective evaluation within real-world EHR workflows.
Background and Aims Detecting subclinical atrial fibrillation (AF) and initiating anticoagulation therapy are critical for secondary stroke prevention after cryptogenic stroke. This study aimed to evaluate the effectiveness of traditional AF risk scores commonly used in clinical practice and compare with an artificial intelligence (AI)-driven ECG-derived AF prediction score. Methods A retrospective tertiary care center study identified all cryptogenic stroke/TIA patients admitted between 2017-2023 who received an implantable cardiac monitor (ICM). The European Society of Cardiology (ESC)-recommended risk factors and risk score components (i.e.,CHA2DS2-VA, Brown ESUS-AF, and HAVOC) from the 2024 guidelines for AF management, were analyzed using multivariate logistic regression to predict AF within one-year post-ICM insertion. The predictive performance of these risk scores and an AI-driven ECG algorithm for AF detection was assessed and compared. Results 230 cryptogenic stroke/TIA patients underwent ICM insertion. Only age (OR: 1.033, 95%CI 1.002-1.066) was associated with an increased likelihood of AF detection within one-year post-ICM insertion. The ECG-AI score demonstrated higher predictive power over CHA2DS2-VA and HAVOC (p=.002), and was the only score exceeding an AUROC value of 0.7, but did not outperform Brown ESUS-AF (p=.210). Patients with high scores had a more than two-fold increased likelihood of AF detection (HR: 2.7, 95%CI 1.2-5.8, p=.01). Conclusions In patients with cryptogenic stroke receiving an ICM, the ECG-AI score showed modest discrimination for AF detection, outperforming CHA2DS2-VA and HAVOC, but not Brown ESUS-AF. Age was the only significant clinical predictor. These findings indicate a possible role for AI-driven ECG analysis in risk stratification.
F. Wouters, M. Barthels, J. Vranken et al.· Digital Health· 0 citations
ABSTRACT Background Cardiac implantable electronic devices (CIEDs) frequently detect brief, often subclinical atrial fibrillation (AF), but their value for predicting progression to persistent AF remains uncertain. Statistical and machine learning (ML) approaches may enable dynamic risk stratification using this longitudinal device data. Objective To develop a risk stratification model using clinical and CIED‐derived AF burden measured over a rolling 6‐month window to predict progression to persistent AF. Methods We analyzed continuous CIED data from 1985 patients without prior persistent AF implanted between 2016 and 2024 at a tertiary medical center. AF burden and clinical variables were summarized using overlapping 6‐month rolling windows to estimate 1‐year risk of persistent AF. Associations were evaluated using Kaplan–Meier and Cox proportional hazards models. A gradient‐boosted decision tree model (XGBoost) was used to predict progression. Results During a mean follow‐up of 1192 days, 874 patients (44%) developed paroxysmal AF, of whom 257 (29%) progressed to persistent AF after a mean of 813 days. Patients with no AF or < 1 h/day of AF in the prior 6 months had > 97% 1‐year freedom from persistent AF, whereas those with > 8 h/day had a 63% progression rate. Higher AF burden was strongly associated with progression (maximum HR 8.66, p < 0.001). The ML model demonstrated high predictive performance (sensitivity 99.4%, specificity 95.7%). Conclusion CIED‐detected AF burden is strongly associated with progression to persistent AF. ML‐based analysis of 6‐month device data enables accurate, point‐in‐time risk stratification to support earlier and more targeted clinical management.
A. Nakonechnyi, Shaul Geliaks, I. Goldenberg et al.· Annals of Noninvasive Electr...· 0 citations
BACKGROUND
Screening for atrial fibrillation (AF) on the basis of AF risk may be more effective. We aimed to develop, externally validate, and prospectively test a machine learning prediction model using electronic health records (EHRs) to guide AF screening.
METHODS
We developed and validated a random forest prediction model for new AF within 6 months, using age, sex, and 10 comorbidities (Future Innovations in Novel Detection of Atrial Fibrillation [FIND-AF] 2.0) in EHRs in the United Kingdom (n=2 081 139), Japan (n=7 795 244), Israel (n=2 166 795), Canada (n=627 919), and China (n=149 145). We conducted a prospective study where participants ≥30 years old without AF and with a CHA2DS2-VASc score ≥2 in men and ≥3 in women, stratified by FIND-AF 2.0 into high and low risk, undertook 4 ECG recordings per day for 3 weeks using a handheld ECG recorder, with a primary outcome of newly diagnosed AF. We estimated stroke risk associated with nonanticoagulated AF in patients with high FIND-AF 2.0 risk in the FinACAF (Finnish Anticoagulation in Atrial Fibrillation) registry of patients with AF (n=229 565).
RESULTS
FIND-AF 2.0 was applicable to all EHRs and showed good to excellent prediction performance (United Kingdom: area under the receiver operating characteristic curve [AUROC], 0.819 [95% CI, 0.809-0.829]; Israel: AUROC, 0.835 [95% CI, 0.828-0.842]; Japan: AUROC, 0.751 [95% CI, 0.745-0.757]; Canada: AUROC, 0.747 [95% CI, 0.741-0.753]; China: AUROC, 0.753 [95% CI, 0.725-0.771]), with AUROC>0.7 in men and women in all cohorts, and improved performance compared with CHA2DS2-VASc and C2HEST. Of 1923 participants from 15 sites in the prospective study (mean age, 70.2 [SD 9.4] years), with a mean of 74.8 (SD, 19.4) ECG recordings, AF was diagnosed in 5 of 902 (0.6%) with low FIND-AF 2.0 risk and 46 of 1021 (4.5%) with high FIND-AF 2.0 risk (odds ratio, 8.46 [95% CI, 3.35-21.40], P<0.001). Median AF burden among high FIND-AF 2.0 risk-detected cases was 33.4% (interquartile range, 5.1%-91.6%), and 96.1% initiated oral anticoagulants. In the FinACAF registry, the rate of ischemic stroke for patients with high FIND-AF 2.0 risk, AF, and no anticoagulants was 6.0 events per 100 patient-years.
CONCLUSIONS
The EHR-based machine learning model, FIND-AF 2.0, identifies a high-risk subpopulation for AF diagnosis among patients at elevated risk of stroke and could enable scalable, EHR-driven, risk-guided AF screening.
R. Nadarajah, Jianhua Wu, A. Wahab et al.· Circulation· 0 citations
Background Australian guidelines recommend the CHA2DS2-VA (congestive heart failure, hypertension, age ≥75 years (double weight), diabetes mellitus, previous stroke (double weight), vascular disease, age 65–74 years) score to identify patients with atrial fibrillation (AF) at low stroke risk who should avoid oral anticoagulation. Treatment thresholds were derived from heterogeneous international data, and alternate scores require validation in an Australian population. We evaluated whether existing stroke scores (CHADS2 (congestive heart failure, hypertension, age ≥75 years, diabetes mellitus, previous stroke (double weight)), CHA2DS2-VA, ATRIA (anticoagulation and risk factors in atrial fibrillation) and Modified-CHADS2) accurately identify truly low-risk patients in Australia. Methods We conducted a population-based cohort study using linked data from the New South Wales Admitted Patient Data Collection, the National Death Index and the Pharmaceutical Benefits Scheme databases. Oral anticoagulant-naïve patients with a first hospital admission for AF between July 2003 and January 2021 were included. Patients were followed for ≥12 months and classified into low-risk, intermediate-risk or high-risk categories for each score. ‘Truly low-risk’ was defined as annual ischaemic stroke incidence <0.9%. Predictive accuracy was assessed using the concordance-statistic. Results Among 224 451 eligible patients, 3552 (1.6%) were hospitalised for ischaemic stroke within 12 months. The proportion labelled low-risk ranged from 2.0% (Modified-CHADS2) to 50.9% (ATRIA). Patients assigned low-risk using each score, and intermediate-risk with the CHA2DS2-VA score, met the definition of truly low-risk. Concordance-statistics were modest, with the weakest performing score being the CHA2DS2-VA score (0.61, 95% CI 0.60 to 0.61) and the best performing model being the ATRIA score (0.66, 95% CI 0.66 to 0.67). Conclusions The guideline-endorsed CHA2DS2-VA score showed the lowest discrimination, and patients classified as intermediate-risk using this score had stroke incidence below the treatment threshold. Strong consideration should be given to the development and validation of locally applicable risk tools.
N. Zubrzycki, K. Hyun, Erdahl Teber et al.· Open Heart· 0 citations
Background Coronary heart disease (CHD) and atrial fibrillation (AF) frequently coexist, yet existing risk stratification tools inadequately capture the nonlinear, multidimensional determinants of AF in middle-aged and older CHD patients. This study aimed to develop and validate an interpretable machine learning-based prediction model leveraging electronic medical records (EMR) data. Methods A retrospective cohort of 47,617 hospitalized CHD patients (January 2020–December 2025) was analyzed. After random forest imputation and least absolute shrinkage and selection operator (LASSO) screening, eight machine learning algorithms were trained and validated (7:3 split). Model performance was evaluated by the area under the receiver operating characteristic curve (AUC), calibration curves, and decision curve analysis (DCA), with Shapley additive explanations (SHAP) and local interpretable model-agnostic explanations (LIME) applied for interpretability. Results LASSO regression identified 16 predictors, with pulse rate, total cholesterol, systolic blood pressure, creatinine, and triglycerides emerging as the top five contributors. XGBoost outperformed competing models, achieving AUCs of 0.867 (95% CI: 0.862–0.872) in the training set and 0.813 (95% CI: 0.802–0.823) in the validation set. Restricted cubic spline analysis revealed nonlinear dose–response relationships for multiple continuous variables. SHAP visualization quantified individualized feature contributions. LIME explanations demonstrated consistent local feature contributions at the individual level. Conclusions This data-driven, interpretable XGBoost model enables individualized AF risk assessment in middle-aged and older CHD patients, offering a practical tool for early identification and targeted intervention in clinical practice.
Feng Chen, Qin Fu, Ling Li et al.· Frontiers in Cardiovascular...· 0 citations
BACKGROUND
Atrial fibrillation (AF) increases ischemic stroke risk nearly fivefold. Although the CHA2DS2-VA score is widely used for risk stratification, its predictive performance is limited, particularly in intermediate-risk patients. Left atrial (LA) enlargement, a marker of structural remodeling, may improve risk prediction.
AIMS
To evaluate whether adding LA diameter to CHA2DS2-VA (creating CHA2DS2-VALa) enhances ischemic stroke prediction in AF patients. We additionally derived a multivariable β-weighted model and benchmarked its performance against CHA2DS2-VA (and the simplified +1 LA rule).
METHODS
This retrospective study included 1648 AF patients. LA diameter was measured by standardized echocardiography. A ROC-derived cut-off (> 4.7 cm) defined enlargement. Patients were followed for a mean of 4.18 ± 0.35 years. Imaging-confirmed ischemic stroke events were identified through institutional hospital records and the electronic health registry, and were validated using hospitalization documentation, neurology consultation notes, discharge summaries, and available brain CT/MRI findings. Predictors were assessed using multivariable logistic regression, and Kaplan-Meier/log-rank analyses compared CHA2DS2-VALa versus CHA2DS2-VA.
RESULTS
Imaging-confirmed ischemic stroke events occurred more frequently in patients with LA > 4.7 cm (25 [3.8%] vs. 21 [2.1%], p < 0.001). Elevated CRP, NT-proBNP, diabetes, hypertension, CAD, and heart failure were also associated with stroke. LA > 4.7 cm independently predicted events (OR 1.97, 95% CI 1.36-2.86, p < 0.001). ROC analysis showed strong discrimination for LA diameter (AUC 0.872; sensitivity 82%, specificity 85%). Kaplan-Meier analysis demonstrated superior prediction with CHA2DS2-VALa compared to CHA2DS2-VA (log-rank p < 0.001).
CONCLUSION
The CHA2DS2-VALa score significantly improves stroke risk stratification in AF by integrating LA diameter into conventional scoring. As echocardiographic measurement is widely available and reproducible, this modification offers a simple, pragmatic approach to refine residual stroke-risk stratification and support closer evaluation and follow-up in anticoagulated AF patients. Prospective validation in diverse populations is warranted.
S. Ömür, Emin Koyun, G. Genc Tapar et al.· Cardiovascular Electrophysio...· 0 citations
A new method for surgically removing training examples from a model reveals that as datasets grow, the link between what a model learns and what it produces dissolves.
