AI Tool Outperforms PD-L1 for Predicting Lung Cancer Immunotherapy Outcomes

An artificial intelligence system developed using data from nearly 2,400 patients with advanced non-small cell lung cancer (NSCLC) outperformed several established biomarkers for predicting outcomes following immunotherapy and improved physicians’ ability to anticipate treatment response.

The findings, published in Nature Medicine, come from I³LUNG, an international effort to develop an AI-based physician decision-support system for immunotherapy in NSCLC. The study included 2,396 patients with stage IIIC to IVB disease treated across six centers in six countries. Researchers incorporated clinical and blood measurements alongside CT imaging, digital pathology, and genomic information into machine-learning and deep-learning models.

Routine clinical data outperform established biomarkers

One of the study’s most notable findings was that relatively simple information already routinely collected in clinical practice carried substantial predictive power.

Models using clinical and blood data, including PD-L1 expression, ECOG performance status, smoking status, metastatic sites, neutrophil-to-lymphocyte ratio, and lactate dehydrogenase, achieved area under the curve values of up to 0.77 in the independent test set.

The AI models significantly outperformed individual predictors including PD-L1, ECOG performance status, neutrophil-to-lymphocyte ratio, lactate dehydrogenase, and the Lung Immune Prognostic Index. PD-L1 remains the principal clinically approved biomarker used to help guide immunotherapy decisions in NSCLC, despite its limited ability to accurately distinguish patients who will experience durable benefit.

Explainable AI improved physician predictions

The researchers also tested whether providing physicians with AI predictions could improve clinical decision-making rather than simply comparing algorithmic performance with conventional biomarkers.

Twenty physicians, including 10 lung cancer experts and 10 nonexperts, assessed real-world patient cases before and after receiving predictions from the model together with explanations showing which features contributed to its conclusions.

For predicting disease control, sensitivity increased from 0.72 without AI assistance to 0.87 with the explainable AI tool, while overall accuracy increased from 0.57 to 0.65. Both lung cancer specialists and nonexperts showed improvements. Agreement between the two groups also increased after AI support was introduced.

The findings suggest that explainability may be important for translating predictive algorithms into clinical tools, allowing physicians to examine the factors underlying an individual prediction instead of receiving an unexplained risk score.

More data did not always mean better predictions

I³LUNG also tested whether combining clinical information with CT imaging, digital pathology, and genomic data could further improve prediction.

Some multimodal machine-learning models showed substantial improvements during cross-validation, particularly in selected patient subgroups. However, those gains were not consistently reproduced in independent testing or external validation. Deep-learning approaches similarly showed no consistent benefit from adding additional modalities.

Performance of the clinical-data models also declined in the geographically distinct external validation cohort, with AUCs ranging from 0.55 to 0.72, highlighting the challenges of transferring AI models between patient populations and healthcare systems.

The study is retrospective, and the researchers caution that the incremental value of multimodal AI remains uncertain. The I³LUNG decision-support system is now undergoing prospective validation in more than 2,000 patients.

If confirmed prospectively, the results suggest that precision immunotherapy may not always require increasingly complex molecular datasets. Combining routinely available patient information with explainable AI could provide a more scalable route toward individualized treatment decisions.