Integrative Identification of Frailty Status in Elderly Patients with Atrial Fibrillation Using a Noise-Resilient...

Frailty is common yet underdiagnosed in elderly patients with atrial fibrillation (AF), worsening outcomes and complicating treatment. Traditional assessments are time-consuming and subjective, while physiological monitoring offers potential for automated detection. Combining machine learning (ML) with multimodal data, especially electrocardiogram (ECG) and clinical features, may improve frailty identification, even in noisy real-world conditions. To develop a noise-resilient ML framework for identifying frailty status in elderly patients with AF using multimodal clinical and ECG data and to compare traditional and deep learning models under varying signal conditions.

This retrospective study included 110 patients aged ≥65 with documented AF. Frailty was assessed via a composite of the Fried Phenotype and Clinical Frailty Scale. Electrocardiogram data (resting 12-lead and Holter) were processed through a denoising and segmentation pipeline. Over 180 ECG features and clinical parameters were used as inputs. Five models (Random Forest, Extreme Gradient Boosting, 1-dimensional convolutional neural network, bidirectional long short-term memory-attention, SiamAF) were trained and evaluated using an 80:20 split. Performance metrics included accuracy, F1-score, receiver operating characteristic-area under the curve (ROC-AUC), and Brier score; robustness was tested with synthetic noise.

Frailty and pre-frailty prevalence were 41.82% and 34.55%. The SiamAF achieved the best performance (accuracy 90.00%, F1-score 89.75%, ROC-AUC 93.00%, Brier score 0.084), maintaining robustness under noise. Deep learning models also showed strong performance (ROC-AUC > 90%). Key predictors included standard deviation of normal-to-normal (NN) intervals, corrected QT interval, N-terminal pro-brain natriuretic peptide, and grip strength.

Multimodal ML effectively identifies frailty status in elderly patients with AF. The SiamAF model demonstrated strong discriminatory performance and noise robustness in this single-center cohort; however, these findings should be considered hypothesis-generating and require external validation in larger, multicenter AF populations before clinical adoption. Given the limited sample size relative to the high-dimensional feature space, these results should be interpreted as exploratory and hypothesis-generating. The present framework primarily leverages predefined ECG-derived features rather than fully end-to-end raw-signal learning.