An international research team has built an AI system that helps doctors more accurately predict whether immunotherapy will control advanced non-small cell lung cancer. In a study involving 20 physicians, the tool lifted accuracy in predicting disease control from 57% to 65%, according to findings published as part of the I3LUNG project.
How the system was trained
Researchers collected data from 2,396 patients across six medical centers in Italy, Germany, Spain, Greece, the United States and Israel. All patients had received immunotherapy, either alone or combined with chemotherapy. The team trained AI models to predict both disease control and patient survival using clinical information and blood test results. More advanced versions of the system also incorporated CT scans, tumor tissue images and genetic data.
The models were designed to identify patterns that current indicators often miss. Doctors today rely on a limited set of biomarkers that do not always provide reliable predictions about who will benefit from immunotherapy.
Testing the tool with practicing physicians
To gauge real-world usefulness, the researchers recruited 20 doctors. Half specialized in lung cancer. The rest were oncologists without a lung cancer focus or physicians still in training. Each doctor first assessed cases independently, then received the AI's predictions along with explanations showing which data points influenced the results.
With AI support, doctors correctly identified 87% of patients whose treatment achieved disease control - tumor shrinkage or stabilization - compared with 72% without assistance. The improvement came with a trade-off. Physicians became slightly more likely to predict a positive outcome for patients whose disease was not actually controlled by treatment.
What comes next
The team is now running further validation studies with more than 2,000 patients. The system's reliability must be confirmed before it can move into clinical practice. Researchers emphasized that the tool is intended to support, not replace, physician judgment.
The study sits within a broader push to apply machine learning to oncology. Similar efforts are advancing across AI for Healthcare and AI for Science & Research, where models trained on multimodal patient data are being tested for diagnostic and prognostic tasks.
Why this matters for science and research professionals
The study illustrates a practical pattern worth watching: AI models trained on routine clinical data - blood work, scans, pathology images - can sharpen human decision-making without requiring exotic new tests. For researchers building or evaluating clinical AI, the 8-percentage-point accuracy gain and the accompanying trade-off in false positives offer a concrete benchmark. The finding that even non-specialist physicians improved with AI support suggests these tools could narrow expertise gaps in settings where lung cancer specialists are scarce.
Your membership also unlocks:
