Artificial Intelligence in Cancer Diagnosis: Relying on Correlated Signals
Study in Nature Biomedical Engineering shows AI models rely on correlated tissue features.
Analysis of 8,000 patient samples found AI accuracy at 80% vs 75% using tumour grade.
Premature adoption risks inappropriate therapies as AI confuses correlated signals.
Models should not replace molecular testing and need stronger evaluation before deployment.
Authors propose moving to causal, biology-aware models with rigorous, bias-aware evaluation.
1 day ago