The European I3LUNG project brought together 2,396 patients with metastatic non-small cell lung cancer to train models combining imaging, pathology and genomics.
Immunotherapy has changed the management of non-small cell lung cancer, but it only benefits 20 to 30% of patients in the long term, and there is currently no marker that makes it possible to designate in advance those who will respond. It is this blind point that I3LUNG tackles, presented as the largest international multimodal real-life study on the subject, with 2,396 patients included, the results of which are published in Nature Medicine.
Researchers fused clinical and blood field data, CT images, digital pathology, and genomics, into early fusion machine learning and intermediate-fusion deep learning models. Performance exceeds that of standard markers — PD-L1 expression, ECOG index, neutrophil-to-lymphocyte ratio, LDH, and LIPI score — with areas under the curve reaching 0.86 in some clinical subgroups.
The most interesting aspect for practice concerns the actual use of the tool: a usability study shows that both specialist and non-specialist doctors improve their predictions when they rely on the explainable version of the system. The tool is presented as a clinically relevant decision support device, now being validated prospectively on more than 2,000 patients.
Read the full article on the original source
Go to the source →