Abstract
Frailty is a critical age-related syndrome associated with increased risks of disability, hospitalization, and mortality. While Comprehensive Geriatric Assessment (CGA) is the standard for evaluation, it is often subjective, time-consuming, and resource-intensive. To address these limitations, this study developed and validated a multimodal deep learning framework that integrates chest X-rays (CXR) and electrocardiograms (ECG) to provide an objective and rapid tool for frailty classification. We retrospectively analyzed data from 2,830 older adults who underwent CGA between 2007 and 2024, partitioned chronologically into training, validation, and test sets. Frailty was defined using a Multidimensional Frailty Score (MFS) > 5. Three deep learning models were trained and evaluated: a CXR-only model (EVA-X vision transformer), an ECG-only model (DenseNet-121), and a multimodal model combining both inputs. The multimodal model outperformed the single-modality models and four non-imaging clinical baseline models incorporating age, sex, comorbidity index, or routine laboratory variables, achieving the highest area under the receiver operating characteristic curve (AUROC) of 0.778 and area under the precision-recall curve (AUPRC) of 0.726. For continuous MFS estimation, the multimodal model achieved the highest Pearson correlation (r = 0.610) and the lowest mean absolute error (MAE = 1.68) among the evaluated models. Gradient-weighted Class Activation Mapping (Grad-CAM) was used to qualitatively explore the image regions that influenced model predictions. This multimodal AI framework demonstrates modest classification performance and feasibility for frailty screening. By leveraging routinely collected clinical images, this framework may serve as a practical screening adjunct to identify older adults who may benefit from formal CGA. Further studies are needed to determine its clinical utility when applied preoperatively for the stratification of postoperative risk.
Funding Information
  • Seoul National University Bundang Hospital (22-2024-0008)
  • Seoul National University Bundang Hospital (09-2024-0002)

This publication has 38 references indexed in Scilit: