Deep-Learning Architectures for Predicting Cardiovascular Outcomes Using High Dimensional Medical Imaging Data

Abstract

This quantitative study examined deep-learning architectures for predicting cardiovascular outcomes using high-dimensional medical imaging data within a retrospective, multi-center observational cohort framework. Imaging examinations were treated as baseline predictors, and clinically documented cardiovascular events were operationalized as outcome labels using predefined event windows and censoring logic. The unit of analysis was the individual patient, and one index imaging examination was retained per patient to ensure independence of observations. The final analytic sample included 1,248 patients drawn consecutively from four clinical sites, with a mean age of 57.6 years (SD = 12.9) and 56.7% male representation (n = 708). Imaging modalities included cardiac MRI (51.0%), CT/CT angiography (26.0%), and cine echocardiography (23.1%). Composite constructs were derived to represent structural imaging risk, functional imaging dynamics, tissue characterization, and clinical risk covariates, with all constructs standardized prior to modeling. Reliability analysis demonstrated satisfactory internal consistency across constructs, with final Cronbach’s alpha values ranging from 0.80 to 0.88. Regression analyses were conducted using stepwise specifications, beginning with clinical controls and expanding to imaging constructs and an integrated imaging–clinical score. Structural imaging risk (β = 0.287, p < .001), functional imaging dynamics (β = 0.214, p = .001), and tissue characterization (β = 0.246, p < .001) were each significantly associated with cardiovascular event occurrence. The integrated imaging–clinical score produced the strongest adjusted association (β = 0.521, 95% CI [0.407, 0.635], p < .001) and yielded the best model fit (AIC reduced from 914.7 to 889.2; pseudo R² increased from 0.167 to 0.193). Model diagnostics indicated acceptable multicollinearity (max VIF = 2.18) and adequate calibration (calibration slope = 0.97; intercept = 0.03). Holdout performance demonstrated strong discrimination (AUC = 0.82) and low probabilistic error (Brier score = 0.098), with specificity of 0.90 and sensitivity of 0.63 at a 0.50 threshold. Robustness checks, including bootstrap resampling and subgroup stratification, produced consistent estimates. Overall, the findings indicated that multi-domain imaging constructs significantly predicted cardiovascular outcomes, and integrated imaging–clinical modeling provided the strongest and most stable predictive evidence.

Ähnliche Arbeiten