Healthcare AI Daily Briefing
New to OpenRounds? Get the free daily email or see how we work.
Medical Imaging
Transformer deep‑learning model matches logistic regression for predicting 6‑month taste alteration after HNC radiotherapy but is less well calibrated.
Journal article (Oral oncology)
In a prospective cohort of 972 head‑and‑neck cancer patients, a 3D transformer model (TransRP) achieved an AUC of 0.71 on a held‑out test set, equal to a conventional NTCP logistic‑regression model (AUC 0.71) and better than a ResNet CNN (AUC 0.68). Calibration was best for the reference logistic‑regression model. Attention maps for the transformer highlighted supraglottic and submandibular regions, and shifting the 3D crop improved deep‑learning performance. [1]
Medical Imaging
Gradient Boosting model using clinical and MRI features differentiates cerebral tuberculoma from lung cancer brain metastasis.
Journal article (Academic radiology)
A Gradient Boosting model built on four predictors — age, CEA level, multiple lesions, and ring‑with‑mural‑nodule appearance — achieved an AUC of 0.970 on an internal test set of 57 patients and 0.957 on an external test set of 51 patients from another hospital. On the external set, sensitivity was 88.0% and specificity 92.3%. Feature selection was restricted to the training/validation split (n=131) and the four predictors were identified via multivariate logistic regression with reported odds ratios. [2]
Diagnostics & Pathology
HRDPath predicts HRD in high‑grade serous ovarian cancer with AUC 0.846 and 0.90 sensitivity and specificity.
Journal article (Nature communications)
HRDPath is a patient‑level deep learning model that aggregates multiple H&E whole‑slide images to predict homologous recombination deficiency without additional molecular testing. Validated across three female cohorts, it attained an AUC of 0.846, significantly outperforming earlier results, and both sensitivity and specificity reached 0.90 through targeted optimization and multi‑slide input. The interpretability analysis showed that the model highlights morphological and pathological changes at the cellular and tissue levels associated with HRD‑positive tumors. [3]
Research & Evidence
Lipid panel plus ApoB modestly predicts first heart failure three years out; high AUCs come from case‑control contrast.
Preprint, not peer reviewed
A preprint using All of Us multicenter EHR data applied machine learning with SHAP to baseline lipid profiles plus apolipoprotein B. In age‑matched men and women the model yielded AUC 0.624 and 0.627 for a first heart‑failure event within 36 months. A separate comparison against age‑matched controls without prior cardiovascular disease produced higher AUCs of 0.937 (men) and 0.925 (women). SHAP analyses identified higher triglycerides, lower HDL and LDL, and ApoB as key predictors, noting LDL’s counter‑intuitive inverse link. [4]
Medical Imaging
TRAM-based deep learning model outperforms DCE-MRI for breast lesion classification.
Journal article (Journal of medical imaging (Bellingham, Wash.))
In a study using the Advanced MRI Breast Lesions dataset from The Cancer Imaging Archive, a 3D ResNet18 model trained on treatment response assessment maps (TRAMs) achieved higher AUC (0.870 vs 0.835), sensitivity (0.848 vs 0.818), and specificity (0.823 vs 0.759) than the same architecture trained on multiphase DCE-MRI. Test-time augmentation estimated aleatoric uncertainty; low-uncertainty TRAM predictions maintained performance while high-uncertainty DCE-MRI predictions declined markedly. Radiomics models showed similar trends. The authors suggest TRAM-based deep learning may provide a complementary imaging approach for breast lesion classification. [5]
Sources
- Predicting taste alteration in head and neck cancer patients using 3D transformer-based deep learning model. · Oral oncology Original source
- Machine Learning Model Using Clinical and Contrast-Enhanced MRI Features Differentiates Cerebral Tuberculoma from Lung Cancer Brain Metastasis. · Academic radiology Original source
- HRDPath: an explainable multi-model deep learning architecture for predicting homologous recombination deficiency from histopathology images. · Nature communications Original source
- Cholesterol Paradox: Multi-point Heart Failure Prediction from Lipid Panels in Electronic Health Records · medRxiv Original source
- Deep learning-based breast lesion classification using treatment response assessment maps with uncertainty quantification. · Journal of medical imaging (Bellingham, Wash.) Original source
Daily healthcare AI brief
Get the free healthcare AI briefing
One concise briefing built from reporting, research, policy, blogs, videos, and full podcast transcripts.
Free. Get a sign-in link by email to finish subscribing. An account is created if needed. Unsubscribe anytime. Already subscribed? Sign in to save and personalize.