Development of an Interpretable Model for Improving Differential Diagnosis in Subjects with a Left Ventricular Ejection Fraction Ranging from 40 to 55%

Distinguishing between Ischemic Heart Disease (IHD) and Non-Ischemic Dilated Cardiomyopathy (DCM) can often be difficult without invasive coronary angiography, especially in patients with Left Ventricular Ejection Fraction (LVEF) ranging from 40 to 50%. Moreover, although it is rare, some healthy subjects (HC) can have an LVEF of about 50% and must be differentiated from IHD and DCM patients. Global longitudinal strain (GLS) and heart rate variability (HRV) analysis are efficient diagnostic tools for different cardiac conditions. The use of interpretable machine-learning methods to direct the diagnosis is also gaining popularity. Therefore, this study aimed to produce a multinomial logistic regression model based on HRV, GLS and clinical features for differential diagnosis between DCM, IHD, and HC in cases with LVEF in a range of 40–55%. The study encompassed 73 DCM, 71 IHD, and 70 HC. The model was produced by logistic regression algorithms considering the set of selected features chosen with the information gain ratio method. The results showed that the most informative features for classification between HC, DCM, and IHD were GLS, meanRR, sex, age, and LFn. The model has a moderately high classification accuracy of 73%. Finally, the developed model with its nomograms enables probabilistic interpretation of classification output between HC, DCM, and IHD, and may support the differential diagnosis in this population.