Artificial Intelligence-Enabled Electrocardiography for Preoperatively Detecting Atrial Fibrillation and Mortality Risk in Patients with Sinus Rhythm

Abstract

Background: Pre-existing atrial fibrillation (AF) and postoperative new-onset AF (NOAF) are independent perioperative risk factors associated with increased short-term mortality and adverse events. This study aimed to develop and validate an artificial intelligence (AI) model capable of detecting hidden AF, including both pre-existing AF and NOAF, from sinus rhythm electrocardiograms, to improve perioperative risks assessment. Methods: We trained and validated an AI model to detect hidden AF. Subsequent analysis confirmed the prognostic relevance of both pre-existing AF and NOAF in patients receiving non-cardiac surgery. The AI model was applied to patients without known AF to evaluate its predictive capability for NOAF and to stratify short-term clinical outcomes. Results: The AI model demonstrated an area under the receiver operating characteristic curve of 0.87 during the development phase for predicting AF. In an independent validation cohort, pre-existing AF and postoperative NOAF were significantly correlated with increased 30-day all-cause mortality. Patients without pre-existing AF who were classified as high-risk by the AI model had substantially higher 30-day all-cause mortality than their low-risk counterparts (HR 17.33, 95% CI 5.29-56.75). Furthermore, the model scores surpassed conventional clinical risk scores in predicting NOAF and 30-day all-cause mortality. Conclusions: This AI-based approach facilitated the accurate identification of patients with elevated perioperative AF-related risk. It will facilitate focused interventions that may enhance clinical outcomes.

Keywords: Mortality Risk; artificial intelligence; atrial fibrillation.

Figure 1

Flow diagram. The artificial intelligence (AI) model was trained using the dataset from Hospital A, with the remaining patients not involved in the training stage of the AI-electrocardiogram (ECG) analysis used for subsequent validation (details are provided in Supplementary Figure S1).

Figure 2

Risk of all-cause mortality after surgery. The AI-ECG-identified high-risk was defined by the cut-off point corresponding to a high positive predictive value (p > 0.994, Figure S3), while the medium-risk was defined by the cut-off point associated with high sensitivity (p > 0.047, Figure S2). (a) A comparison of patients with and without observed atrial fibrillation, involving 13,678 patients to validate a previous study. The hazard ratio (HR) was adjusted using inverse probability weighting of the propensity score (IPWPS). (b) The relationship between AI-ECG prediction and all-cause mortality in patients without a history of atrial fibrillation, including 13,580 patients, demonstrating the benefits of AI-ECG. HRs were adjusted for age and sex. (c) Stratified analysis of observed atrial fibrillation, with HRs adjusted for age and sex. Abbreviations: hx, history; AF, atrial fibrillation; HR, hazard ratio.

Figure 3

Secondary clinical outcomes stratified by AI-ECG risk categories. Individuals classified as high risk for hidden AF exhibited the highest cumulative incidence rates of new-onset ischemic stroke (HRs of 18.06; 95% CI 5.65 - 57.78), acute myocardial infarction (HRs of 36.06; 95% CI 5.51 - 236), and heart failure (HRs of 12.67; 95% CI 3.45 - 46.52) within 30 days post-surgery, followed by those at medium and low risk. The comparison of patients across different risk categories based on AI-ECG predictions is summarized in the right panel.

Figure 4

Comparison of AI-ECG and clinical risk scores on identifying new-onset atrial fibrillation. Analyses included the data of 13,580 patients without pre-existing atrial fibrillation to validate AI-ECG benefits. (a) Receiver operating characteristic curve analysis of postoperative atrial fibrillation within 1 month, with cut-off points for AI-ECG, Taiwan AF scores, and C2HEST set at 0.047 (defined in Figure S2), 5.5, and 2.5, respectively. (b) One-month follow-up analyses, where the AI-ECG-identified high risk was defined by a high positive predictive value cut-off (p > 0.994, Figure S3) and the medium-risk by a high sensitivity cut-off (p > 0.047, Figure S2). C-indexes were calculated using continuous scores. Abbreviations: AF, atrial fibrillation; HR, hazard ratio.