Machine learning and deep learning predictive models for prognosis in patients with atrial fibrillation: a systematic...

To summarise the performance of machine learning (ML) and deep learning (DL) prognostic models for atrial fibrillation (AF), compare their relative performances with non-artificial intelligence (AI) methods, and to identify key research gaps.

We searched PubMed, Embase, Scopus, the Cochrane Library, Web of Science, and ProQuest from inception to 21 October 2024 for cohort, case-control, cross-sectional, and randomised controlled studies that used ML or DL models to predict clinical outcomes in AF patients. Studies were excluded if they focused on non-AF populations, lacked model performance evaluation, or were abstracts, reviews, or other non-primary research articles. Extracted information included study characteristics, patient demographics, model details and validation strategies. Reporting quality and risk of bias were assessed using the TRIPOD+AI (Transparent Reporting of a multivariable prediction model for Individual Prognosis Or Diagnosis for AI) and PROBAST+AI (Prediction model Risk Of Bias ASsessment Tool for AI) checklists. The primary outcome was model discrimination, measured by the area under the receiver operating characteristic curve (AUC). Meta-analyses were conducted, with heterogeneity assessed via Cochran's Q test and I² statistics.

Of the 7128 studies identified, 81 fulfilled the selection criteria. Among these, 57 applied ML models (81 models total) and 24 used DL models (31 models total). Commonly predicted outcomes included AF recurrence (n=43), ischaemic stroke (n=20), all-cause mortality (n=15), major bleeding (n=11), heart failure (n=3), major adverse cardiovascular events (MACE) (n=3) and thromboembolic events (n=3). AI models exhibited moderate-to-good predictive performance, ranging from a pooled AUC of 0.71 (95% CI 0.66 to 0.76) for major bleeding to 0.85 (95% CI 0.79 to 0.92) for heart failure. Significant heterogeneity was observed across studies (I² 87%-100%). When evaluated on the same datasets, both AI model types generally outperformed risk scores and regression-based models. PROBAST+AI assessment identified high risk of bias in 66 studies (81%) for model development and 68 studies (84%) for model evaluation, primarily due to inadequate handling of missing data and underpowered datasets.

AI models show great promise in AF prognosis tasks and generally outperform non-AI prediction methods. The substantial heterogeneity limits the clinical interpretability of pooled AUCs and warrants cautious interpretation. Standardised reporting and multimodal data integration will be essential to improving model reliability and clinical applicability of AI prognostic models for AF.

CRD42024606885.