An autopsy-based cardiac lesion evaluation system facilitates quantitative diagnosis of sudden cardiac death: development and multicenter validation of a machine learning model

Abstract Background The clinical prevention and forensic diagnosis of sudden cardiac death (SCD) remain challenging due to the absence of standardized quantitative criteria for evaluating cardiac morphological and histopathological alterations. We aimed to develop a machine learning-driven cardiac lesion evaluation system to facilitate quantitative diagnosis and risk stratification of SCD. Methods A total of 2284 adult autopsy cases from the forensic center at Sun Yat-sen University and 1883 external cases from five independent centers were enrolled to develop and validate an autopsy-based diagnostic model. Eight machine learning algorithms were employed, with the optimal model further tested in human–machine collaborative experiments. The model was subsequently transformed to identify myocardial infarction in a prospective clinical cohort of 204 patients presenting with chest pain. Results SCD cases exhibited significantly greater right ventricular wall thickness (OR: 1.17 [95% CI: 1.04–1.32] per mm) and larger valve annulus circumferences, including tricuspid (OR: 1.17 [95% CI: 1.04–1.33] per cm), pulmonary (OR: 1.54 [95% CI: 1.34–1.76]), mitral (OR: 1.16 [95% CI: 1.04–1.29]), and aortic (OR: 1.24 [95% CI: 1.06–1.44]) valves. The logistic regression model demonstrated strong discriminatory performance for SCD, achieving an area under the receiver-operating characteristic curve (AUC) of 0.839 (95% CI: 0.821–0.858) in the training set and 0.840–0.907 across external validation cohorts. Pathologists assisted by the model showed improved diagnostic accuracy, with higher AUC ( P = 0.004) and sensitivity ( P = 0.01) for SCD diagnosis. For sudden coronary artery death, the morphology-based model achieved an AUC of 0.781 (95% CI: 0.738–0.825), while its performance in detecting myocardial infarction using echocardiography-measurable features yielded an AUC of 0.697 (95% CI: 0.587–0.817). Conclusions This rigorously validated model serves as a novel assistant tool for pathologists to achieve quantitative diagnosis of SCD and provides clinicians with a potential tool for myocardial infarction identification and SCD warning. Graphical Abstract