Machine Learning, Radiomics, and Clinical Models Predicting Checkpoint Inhibitor Pneumonitis in Lung Cancer: A Systematic Review and Meta-analysis.

In 30 seconds

This systematic review and meta-analysis assessed machine-learning, radiomics, deep-learning, and clinical models for predicting checkpoint inhibitor pneumonitis in people with lung cancer. Twelve models, covering 4,875 patients and 1,057 events, contributed to the main analysis. The pooled validation C-statistic was 0.82 (95% confidence interval 0.75-0.88), but heterogeneity was high (I 2 = 90.7%) and the 95% prediction interval was wide (0.51-0.95). The authors judged current models not ready for clinical use, citing low-certainty evidence and limited external validation.

Key findings

• Across validation studies, the pooled C-statistic was 0.82 (95% confidence interval 0.75-0.88); heterogeneity was high (I 2 = 90.7%).
• The 95% prediction interval was 0.51-0.95, indicating substantial variation in expected performance across settings.
• Imaging-based models had numerically higher discrimination than clinical models (0.85 versus 0.78), but the difference was not statistically significant (moderator p = 0.46).
• Only two studies reported external validation, with C-statistics of 0.84 and 0.92; overall risk of bias was high in two studies, unclear in nine, and low in one.

Why it matters

CIP can be potentially fatal, so reliable risk prediction could help identify patients who warrant closer assessment. However, the pooled discrimination should not be interpreted as dependable performance in a new clinical setting: the prediction interval was wide, evidence certainty was low, and external validation was sparse. These findings highlight the need to assess model performance and calibration across independent populations before relying on predictions.

What to keep in mind

The evidence was low certainty, heterogeneity was high, and only two studies reported external validation. Risk of bias was unclear or high in most included studies.