AI Predicts Immunotherapy Benefit More Accurately Than Existing Measures

Artificial intelligence (AI) was found to identify patients likely to benefit from immunotherapy more accurately than existing measures.

According to BioOn on the 28th (local time), an international research team built an AI model integrating clinical and blood data, computed tomography (CT) images, digital pathology slides and genomic information from patients with non-small cell lung cancer, then tested its predictive performance.

The results were published in the international journal Nature Medicine.

Lung cancer is the leading cause of cancer death, claiming about 1.8 million lives each year, and non-small cell lung cancer accounts for more than 80% of cases. Immunotherapy provides long-term survival benefits to 20~30% of patients, but the remaining patients do not respond or are left with the burden of toxicity and cost.

Treatment decisions currently rely heavily on PD-L1 expression levels, but limitations have been noted. In response, the researchers analyzed data from 2,396 patients with advanced non-small cell lung cancer across six countries through the I3LUNG project, the largest multicenter study of its kind globally.

An AI model using only clinical and blood data achieved an area under the curve (AUC) of up to 0.77 in independent testing. AUC is a measure of classification accuracy, with 0.8~0.9 considered excellent.

These models outperformed all existing standard measures, including PD-L1, ECOG performance status, the neutrophil-to-lymphocyte ratio, lactate dehydrogenase and the lung immune prognostic index. Adding imaging and pathology data raised the AUC to 0.88.

However, the researchers said the additional benefit of multimodal integration was not consistent across the test set and external validation. In external validation, the AUC fell to 0.55~0.72; the researchers attributed this to differences between populations.

The benefits of physician-AI collaboration were also confirmed. After 20 physicians independently assessed cases involving 100 patients and then reassessed them with AI assistance, sensitivity for identifying immunotherapy responders increased from 0.72 to 0.87.

The improvement was particularly pronounced among physicians who were not lung cancer specialists. The researchers said the approach could be useful in community care settings where thoracic specialists are scarce.

The findings reported here are from a retrospective phase, and I3LUNG is currently conducting a prospective study involving more than 2,000 people at six centers. The researchers said, “Decision-support tools that outperform existing biomarkers can be developed using only routinely available clinical data.”