Radiomics-based machine learning on noncontrast postmortem CT showed good diagnostic performance for myocardial infarction detection (mean AUROC 0.814–0.817), emphasizing its potential as an objective, noninvasive tool for cause-of-death assessment.
Logistic regression with no additional feature selection achieved the highest diagnostic performance for MI detection on noncontrast PMCT.
Support vector machines and random forest classifiers with no additional feature selection showed comparable diagnostic performance to logistic regression.
Random forest-based and recursive feature elimination with logistic regression feature selection methods did not improve diagnostic performance compared to no additional feature selection across most classifiers.
Four radiomic features were consistently identified by both RF-based ranking and RFE-LR feature selection methods.
The study used a retrospective design with 106 PMCT examinations, with left ventricular myocardium segmented semiautomatically for radiomic feature extraction.
The authors identified potential translational relevance of PMCT radiomics findings for clinical contrast-free cardiac imaging.
This research suggests that artificial intelligence applied to routine postmortem CT scans — the type taken without contrast dye injection — can detect whether a person died from a heart attack with meaningful accuracy. The researchers analyzed 106 CT scans from deceased individuals, half of whom had autopsy-confirmed heart attacks, and used a technique called radiomics to extract hundreds of subtle numerical features from the heart muscle tissue. These features were then fed into machine-learning models, with the best model achieving an area under the ROC curve (a measure of diagnostic accuracy) of about 0.82, meaning it could correctly distinguish heart attack cases from non-heart attack cases roughly 82% of the time on average. The study found that simpler machine-learning approaches — specifically logistic regression without elaborate feature pre-selection — performed just as well as or better than more complex methods. Four specific radiomic features were consistently identified as important across different analytical approaches, though the paper does not detail what tissue characteristics these features represent. The findings are relevant for forensic medicine and pathology, where determining the cause of death without a full autopsy can be challenging, and where postmortem CT scans are already routinely performed in many settings. This research suggests that analyzing CT scans with radiomics and machine learning could one day help medical examiners and forensic pathologists identify heart attack as a cause of death more objectively and without invasive procedures. The authors also note that because the CT scans used were taken without contrast agents — similar to standard clinical cardiac scans — there may be future applications in living patients as well. However, the study was relatively small and retrospective, and the authors emphasize that further validation in larger and more diverse populations is needed before these methods could be used in practice.
Barman N, Rastogi A, Radbruch A, Luetkens J, Wittschieber D, Hahnemann M. (2026). Detection of myocardial infarction by postmortem CT combined with radiomics-based machine-learning analysis.. European radiology experimental. https://doi.org/10.1186/s41747-026-00800-4