Interpretable machine learning for predicting isolated basal septal hypertrophy

• Published in: PloS one Vol. 20; no. 6; p. e0325992
• Main Authors: Gao, Lei, Tian, Boyan, Jia, Qiqi, He, Xingyu, Zhao, Guannan, Wang, Yueheng
• Format: Journal Article
• Language: English
• Published: United States Public Library of Science 30.06.2025; Public Library of Science (PLoS)
• Subjects: Aged; Aged, 80 and over; Algorithms; Aorta; Aortic valve; Biology and Life Sciences; Blood pressure; Body mass index; Cardiac arrhythmia; Cardiomyopathy; Complications; Computer and Information Sciences; Coronary vessels; Data collection; Decision trees; Diagnosis; Disease; Dyskinesia; Echocardiography; Effectiveness; Feature selection; Female; Flow velocity; Health aspects; Heart; Heart enlargement; Heart Septum - diagnostic imaging; Heart Septum - pathology; Heart valves; Hospitals; Humans; Hypertension; Hypertrophy; Learning algorithms; Machine Learning; Male; Medicine and Health Sciences; Missing data; Mitral valve; Patients; Physical Sciences; Postoperative; Prediction models; Regression analysis; Regression models; Research and Analysis Methods; Risk factors; Risk management; ROC Curve; Segments; Septum; Structure-function relationships; Test sets; Ultrasonic imaging; Variables; Ventricle; Ventricular Septum - diagnostic imaging; Ventricular Septum - pathology; China
• ISSN: 1932-6203; 1932-6203
• DOI: 10.1371/journal.pone.0325992

The basal septal hypertrophy(BSH) is an often under-recognized morphological change in the left ventricle. This is a common echocardiographic finding with a prevalence of approximately 7-20%, which may indicate early structural and functional remodeling of the left ventricle in certain pathologies. It also poses a risk of severe left ventricular outflow tract obstruction and is a significant cause of postoperative complications in patients undergoing transcatheter aortic valve implantation (TAVI). Compared to traditional algorithms, machine learning algorithms are more effective at capturing nonlinear relationships and developing more accurate diagnostic and predictive models. However, no predictive models for BSH have been developed using machine learning algorithms. To evaluate the effectiveness of five machine learning algorithms in predicting thickening of the basal segment of the interventricular septum and to develop a simple, yet efficient, prediction model for BSH. Echocardiographic and clinical data from 902 patients were collected from the First Central Hospital of Baoding City, including 91 BSH patients and 811 non-BSH patients. The data were divided into training and test sets in a 7:3 ratio. Five machine learning algorithms -XGBoost, Random Forest(RF), Dicision tree(DT), K-Nearest Neighbor classification(KNN), and Naive Bayes(NB) were applied to construct the models, combined with logistic regression (LR) based on Lasso regression. The performance of each model was evaluated using Receiver Operating Characteristic curve (ROC),calibration curves and Decision Curve Analysis (DCA)curve, with the model demonstrating the best performance being selected. The shapley additive explanation (SHAP) method was employed to interpret the XBoost and RF models. The logistic regression (LR) of the Lasso regression model showed that IVS-AO Angle, Left Ventricular Mass Index (LVMI), Diastolic Left Ventricular Internal Diameter Index (LVIDdI), Systolic Blood Pressure (SBP), Diastolic Blood Pressure (DBP), Distance from mitral valve closure point to basal segment of interventricular septum (MVCP-Sd), GLU, and Mitral Valve peak A (MV-A) were associated with BSH, with odds ratios (OR) of 0.86 (0.831-0.888), 1.034 (1.018-1.052), 0.104 (0.023-0.403), 1.041 (1.021-1.064), 0.964 (0.93-0.998), 0.852 (0.764-0.949), 1.146 (1.023-1.281), and 0.967 (0.947-0.987), respectively. The area under the ROC curve (AUC) for Model-relevant variable IVS-AO Angle, MVCP_Sd,LVMI, GLU, LVIDdI, SBP,DBP,LVIDdI,MV_A were 0.87,0.68,0.66,0.55,0.56,0.67,0.75,0.75. The AUC for the algorithms (XGBoost, RF, DT, KNN, NB) in the test set were 0.92, 0.91, 0.85, 0.84, and 0.88, respectively. The SHAP method identified eight predictor variables for BSH based on importance rankings, with the top four being IVS-AO Angle, LVMI, LVIDdI, and SBP, with IVS-AO Angle emerging as the most important predictor. The external validation of the RF model yielded an AUC of 0.86. Machine learning can effectively predict BSH, with IVS-AO Angle identified as an independent predictor. The RF model, being simple to operate, can be applied to the risk management of BSH patients.