AI on breast biopsy slides flagged intermediate-risk patients who gained from chemotherapy

Applied without retraining to 6,735 TAILORx trial participants, an AI model reading H&E slides alongside clinical variables predicted disease-free interval (C-index 0.736) and, within the intermediate-risk group, separated patients who gained from chemotherapy from those who did not.

A prespecified analysis of the TAILORx randomised trial applied an AI model reading H&E pathology slides alongside clinical variables to 6,735 node-negative, hormone receptor-positive breast cancer patients, with no TAILORx data used in training. The score was associated with disease-free interval in both treatment arms and remained significant after adjustment for the 21-gene Recurrence Score, and within the intermediate-risk group it identified a chemotherapy benefit (5 percentage points higher 5-year disease-free interval) that the genomic assay itself did not detect (interaction p=0.001 for the AI score against p=0.20 for the genomic score).

Why it is interesting: An externally validated, prespecified test of image-based AI as a predictive rather than merely prognostic marker, in a randomised trial where the genomic assay itself found no such interaction.

Source: medRxiv, 20 September 2026
DOI: 10.64898/2026.09.12.26362500
Type: Preprint
Design: Prespecified secondary analysis of the TAILORx phase 3 randomised trial, 6,735 patients with H&E slides and 21-gene Recurrence Score, using an AI model not trained on trial data
Verdict: Refines prior evidence
Driver: AI in medicine
Driver: Cancer

Remote exercise rehabilitation raised peak oxygen uptake in cardiac patients

A meta-analysis of 13 randomised trials in 958 adults with cardiovascular disease found exercise-based telerehabilitation improved peak oxygen uptake by 2.58 mL/kg/min against usual care, on a 95% prediction interval running from -1.28 to 6.44 and substantial heterogeneity.

Emergent technology Medicine out of hospital Cardiovascular disease Diet and physical activity

Item confirms peer-reviewed 23 Sep 2026

AI models predict death after hip fracture but are seldom externally validated

A systematic review of 40 studies published between 2010 and 2026 found that machine learning models for outcomes after hip fracture, most often death, achieved moderate to strong discrimination but were limited by inconsistent external validation and sparse calibration.

Emergent technology AI in medicine World

Item new preprint 21 Sep 2026

Peer-derived fMRI neurofeedback during a film reduced craving in opioid use disorder

A real-time fMRI study trained people with opioid use disorder to align prefrontal activity with a signal derived from peers further into recovery watching the same film, and reported reduced craving against sham feedback, in a preprint that gives no sample size.

Emergent technology AI in medicine Alcohol, drugs and suicide

Item confirms preprint 21 Sep 2026

Gradient boosting did no better than logistic regression on HPV vaccine hesitancy

In 1,156 Cameroonian parents of children aged 9 to 18, XGBoost, random forest and logistic regression predicted HPV vaccine hesitancy almost identically (AUC 0.841-0.870, XGBoost 0.870 against 0.869, p=0.965), with logistic regression the best calibrated.

Emergent technology AI in medicine Infectious disease and pandemics Cameroon

Cite this item