A team at Beijing Chaoyang Hospital builds and compares five supervised machine learning algorithms using clinical, echocardiographic, and hemodynamic features. They identify six key predictors via LASSO, train models with logistic regression, SVM, random forest, XGBoost, and decision tree, and use SHAP to interpret the best model’s decisions in predicting BPA outcomes.

Key points

• Six predictors selected by LASSO: occlusive lesion proportion, TAPSE/PASP, 6MWD, RVESA, TR severity, PVR.
• Logistic regression with L2 regularisation outperforms other ML models, achieving test AUC of 0.865, accuracy 0.848, sensitivity 0.950.
• SHAP analysis identifies occlusive lesion proportion as the most influential feature driving BPA response predictions.

Why it matters: A reliable ML tool for preoperative BPA response prediction can enhance patient selection, reduce procedural risks, and improve outcomes in CTEPH management.

BMC Medical Imaging investigators implement a radiomics pipeline extracting high-order texture features from NCCT scans, co-registered with diffusion-weighted MRI, to train a random forest classifier that accurately discriminates acute ischemic stroke lesions within six hours, facilitating rapid, accessible early diagnosis.

Key points

• Co-registered NCCT and DWI images from 228 acute ischemic stroke patients enable precise infarct labeling for radiomic analysis.
• Ten RPT-selected radiomic features—including wavelet, LoG, and gradient textures—are normalized and input into a random forest classifier.
• Model achieves AUROCs of 0.858/0.829/0.789 and accuracies up to 79.4%, enabling subvisual infarct detection within six hours on standard CT.

Why it matters: Subvisual stroke lesion detection on routine CT scans expedites early intervention and democratizes acute ischemic stroke diagnosis in resource-limited settings.