A standardized benchmarking framework reveals that most cancer drug response prediction models perform far worse than published results suggest, often failing to beat naive predictors or generalize to unseen drugs. Original paper: Critical evaluation of drug response prediction models with DrEval. — Nature communications. 10.1038/s41467-026-72903-w 📄 Read the article

A Norwegian study shows artificial intelligence can match radiologist performance in detecting breast cancer on digital tomosynthesis scans, potentially transforming screening workflows. Original paper: Artificial Intelligence for Digital Breast Tomosynthesis Screening with and without prior Examinations in BreastScreen Norway. — Radiology. Artificial intelligence. 10.1148/ryai.250988 📄 Read the article

Researchers developed a machine learning model that detects peripheral artery disease (PAD) by analyzing photoplethysmography (PPG) signals from toe pulses. The approach achieved 84.5% accuracy when combined with clinical data and demonstrated equitable performance across demographic groups. Original paper: Development and validation of a digital biomarker for peripheral artery disease. — NPJ digital medicine. 10.1038/s41746-026-02655-w 📄 Read the article

A CT-based radiomics model successfully identifies intratumoral fibrosis grade in hepatocellular carcinoma and predicts treatment response to TACE, offering clinicians a noninvasive tool for patient stratification. Original paper: CT Radiomics-based Machine Learning to Identify Intratumoral Fibrosis and Underlying Angiogenesis in Hepatocellular Carcinoma. — Radiology. 10.1148/radiol.250736 📄 Read the article

An AI model can predict 10-year ischemic stroke risk from electrocardiograms with performance comparable to established clinical scores, suggesting atrial electrical dysfunction drives stroke risk in certain patients. Original paper: ECG Signatures and Long-Term Ischemic Stroke Risk: A Deep Learning Analysis of 200,000 Patients. — Journal of the American College of Cardiology. 10.1016/j.jacc.2026.03.084 📄 Read the article

A physics-informed neural network on smartwatches can continuously monitor blood pressure from electrical signals with clinical accuracy, without a cuff. Original paper: Cuffless hemodynamic monitoring with physics-informed machine learning models. — Nature communications. 10.1038/s41467-026-72693-1 📄 Read the article

A conversational AI system that interprets medical images and documents outperformed primary care physicians on most evaluation metrics in simulated clinical consultations, suggesting potential for supporting remote healthcare. Original paper: Advancing conversational diagnostic AI with multimodal reasoning. — Nature medicine. 10.1038/s41591-026-04371-0 📄 Read the article

A machine learning model for predicting kidney graft loss showed no clinical benefit when integrated into routine care, revealing a gap between model performance and real-world practice. Original paper: Randomized trial of electronic health record implemented AI risk prediction in kidney transplant care. — NPJ digital medicine. 10.1038/s41746-026-02757-5 📄 Read the article

DKFusion, a new AI framework, automatically detects diagnoses that were clinically documented but omitted from discharge summaries, improving medical coding accuracy and insurance reimbursement. Original paper: A data and knowledge cross-level fusion-driven learning framework for detecting missing diagnosis. — NPJ digital medicine. 10.1038/s41746-026-02725-z 📄 Read the article

Researchers have developed CrossPET-Adapt, a machine learning method that eliminates CT scans from PET imaging while adapting to new clinical environments with minimal training data. Original paper: Generalizable CT-free PET attenuation and scatter correction via few-shot cross domain adaptation. — NPJ digital medicine. 10.1038/s41746-026-02760-w 📄 Read the article