Radiomics-based classifier predicted ischemic, hypertensive, and tachycardia-induced LV systolic dysfunction with AUCs of 0.83, 0.79, and 0.84, respectively.
Does a radiomics-based classifier using CMR predict the underlying etiology of left ventricular systolic dysfunction in patients with heart failure?
Machine learning using CMR radiomic features and clinical covariates can predict the clinical history underlying a patient's LVSD with statistically significant accuracy.
Absolute Event Rate: 0% vs 0%
Abstract Introduction Heart failure (HF) is a morbid, complex syndrome affecting over 6 million Americans. Approximately half of HF patients have left ventricular systolic dysfunction (LVSD) — an ejection fraction 50%. This phenotype is termed HF with reduced ejection fraction (HFrEF). Determining which etiolog(ies) underly LVSD in HFrEF (e.g., ischemia, tachycardia, hypertension, etc.) is central to guiding management. Cardiac MRI (CMR) is the gold-standard imaging technique for HF and can help to establish LVSD etiology. However, broader CMR adoption is limited by expert reader availability. Additionally, subtle findings that suggest specific etiologies may be imperceptible to human readers, causing diagnostic ambiguity absent clinical corroboration. Radiomics applies computer vision to medical imaging diagnostics. The tool has been used in CMR applications like distinguishing myopathy genotypes. Our research question is: Can a radiomics-based classifier predict the etiolog(ies) underlying a patient’s LVSD using CMR? Methods In this cross-sectional, retrospective study, the UK Biobank (UKBB) was queried for patients with LVSD and diagnostic (ICD-10) codes suggestive of either ischemic (ICM), tachycardia-induced, and/or hypertensive heart disease (HHD)-related pathology. Cine and T1 map images were algorithmically segmented. A supervised feature selection and classification pipeline modeled etiologies as predictions from image features and covariates. An 80%, cross-validated sample split trained the model. A reserved 20% split validated model performance. Results Inclusion criteria identified 476 patients. Internal validation returned binary areas under curves (AUCs) of 0.83, 0.79, and 0.84 for ICM, HHD, and tachycardia-induced LVSD. All AUCs were significant at p0.05. Positive likelihood ratios for a model diagnosis were 3, 3.3, and 2.85 for these three etiologies. Calibration and sensitivity were noted to be poorer for HHD than for other etiologies. Conclusion We demonstrate that machine learning using radiomic features and clinical covariates can predict the clinical history underlying a patient’s LVSD with statistically significant, though imperfect, accuracy. This extends prior literature by applying cardiac radiomics to a subjective, nonbinary, clinical target. Limitations of this study include HHD classifier performance, a lack of external validation, and insufficient sample depth to study rarer, infiltrative etiologies. Such etiologies could represent a future direction.
Stewart et al. (Sat,) reported a other. Radiomics-based classifier predicted ischemic, hypertensive, and tachycardia-induced LV systolic dysfunction with AUCs of 0.83, 0.79, and 0.84, respectively.