A systematic review and meta-analysis of 17 studies covering 44 machine learning models for predicting neoadjuvant immunotherapy response in resectable Non-Small Cell Lung Cancer (NSCLC) found that all models had high risk of bias, even though internal validation performance was moderately strong (pooled AUC 0.786; external AUC 0.760). In the study, 89% of models were rated relatively low quality, with common flaws in sample size, missing-data handling, and validation methods.

Why It Matters To Your Practice

• These tools may look promising on paper, but weak methodology can inflate performance and limit bedside reliability.

• If you are evaluating AI for neoadjuvant decision support, model governance matters as much as AUC.

• The review suggests current evidence is not yet strong enough to support uncritical adoption in routine surgical or perioperative oncology workflows.

Clinical Implications

• Internal validation showed pooled sensitivity of 0.763 and specificity of 0.908, but these estimates came from a limited subset of models.

• External validation performance was lower, with pooled AUC 0.760, underscoring the risk of performance drop-off outside the development setting.

• SVM had the highest reported AUC (0.841), and non-radiomics models outperformed radiomics models (0.869 vs. 0.775).

• Predicted AUC was 0.805 for major pathologic response and 0.761 for pathologic complete response.

Insights

• The biggest recurring problems were unreasonable sample size, improper handling of missing data, and flawed validation procedures.

• Radiomics reporting quality was inconsistent: among 12 radiomics studies, mean RQS was 14.58, indicating substantial reporting and design gaps.

• Applicability was rated good overall, suggesting the main barrier is not clinical relevance but trustworthiness of model development.

The Bottom Line

• AI models for predicting neoadjuvant immunotherapy response in resectable NSCLC show signal, but the evidence base is methodologically fragile.

• For clinicians interested in how AI will affect practice, this review is a reminder to ask not just how accurate a model is, but how it was built, validated, and tested across settings.

• Until stronger external validation and better study design emerge, these models are best viewed as investigational rather than practice-changing.