A systematic review of 35 studies in Breast Cancer found MRI radiomics models predicted pathologic complete response to neoadjuvant therapy with a pooled AUC of 0.813, but 86% of studies were underpowered and only 14% tested performance in clinically relevant subgroups. The PROSPERO-registered review suggests technical promise, but limited external validation and inconsistent demographic reporting remain major barriers to clinical use.

Why It Matters To Your Practice

• pCR prediction could help clinicians better tailor neoadjuvant treatment, surgical planning, and patient counseling.

• AI tools may appear accurate on paper, but weak validation and poor subgroup reporting can limit reliability at the bedside.

• For clinicians evaluating new AI products, this review highlights the gap between model performance and practice readiness.

Clinical Implications

• Reported model discrimination was favorable overall, but most studies were retrospective, single-center, and methodologically heterogeneous.

• External validation was uncommon, so performance may drop when models are applied outside the original development setting.

• Only a small minority of studies assessed subgroup performance, raising concerns about whether predictions generalize across different patient populations.

Insights

• The review assessed 35 studies with sample sizes ranging from 55 to 442.

• Methodological quality was evaluated with the METRICS score, and bias/fairness review was adapted from QUADAS-2.

• Inconsistent reporting of demographic and clinical variables made it difficult to judge generalizability and equitable performance.

The Bottom Line

• MRI radiomics for pCR prediction is promising, but not yet ready for routine clinical deployment.

• Before adoption, clinicians should look for prospective multicenter validation, adequate sample size, transparent reporting, and subgroup-specific performance data.

• In AI for practice, accuracy alone is not enough; trust depends on robustness, reproducibility, and fairness.