ÌÇÐÄTV

We are delighted to showcase another Pathology News podcast featuring our very own Fayyaz Minhaz, Deputy Director of the TIA Centre and lead researcher for PRISM (Predictive Systems in Biomedicine).

This podcast episode focusses on a recent published paper "Confounding factors and biases abound when predicting molecular biomarkers from histological images" in Nature Biomedical Engineering which examined the reliability of AI models that predict molecular biomarkers directly from routine pathology slides.

Fayyaz Minhas and other researchers involved in the work, Nasir Rajpoot and Muhammed Dawood, found that, across multiple cancer types and machine learning architectures, many current models achieve high headline accuracy, but often rely on correlations between biomarkers or clinicopathological features rather than isolating biomarker specific signals. When evaluated carefully across patient subgroups, performance can drop substantially.

The key message is not that AI in pathology is ineffective. Rather, it is that evaluation standards need to mature. Aggregate accuracy alone is not enough. Stratified analysis, bias aware testing, and comparison against simple clinical baselines are essential before considering clinical deployment.

AI can play an important role in research, triage, and hypothesis generation. But if we want these tools to support patient care responsibly, we must ensure they are learning biology, not shortcuts.

Listen to the Podcast .