AI model built from routine clinical data helped physicians predict lung cancer control on immunotherapy more accurately

AI model built from routine clinical data helped physicians predict lung cancer control on immunotherapy more accurately

On 13 September, Nature Medicine published a study from the I3LUNG project covering 2,396 patients across six clinical centers. The authors tested how an AI prediction based on data a physician normally collects before treatment compares with standard biomarkers, and whether it changes clinical assessments of completed cases.

Immunotherapy enables the immune system to attack the tumor. In advanced non-small-cell lung cancer, it is difficult for the treating physician to predict at the start of therapy which patients will have their disease held in check and which will live longer. One reference point is the level of the PD-L1 protein in the tumor; it helps guide treatment selection but captures only part of the patient's situation.

The authors investigated how much prognostic information already exists in a standard medical record. For the primary model, they selected nine features available before therapy: sex, smoking history, the patient's ability to maintain daily activity, PD-L1 expression, metastatic sites, and blood test values. In an independent patient cohort, the model outperformed individual biomarkers and the LIPI index (which is based on blood tests) in distinguishing disease control and several survival outcomes.

CT scans and digital images of tumor sections improved performance when tested on data from the same centers. In independent patient cohorts, this additional benefit appeared inconsistently, while the model relying on clinical data and blood tests proved more robust. In this study, the most reliable prediction came from the set of data a physician obtains before treatment begins.

The authors also tested whether a physician could use such a prediction in practice. Twenty physicians reviewed 100 case histories: first using patient data alone, then with the model's output and an explanation of which features shifted the prediction in each specific case. After receiving the model's guidance, physicians more often recognized cases in which disease control was achieved: the rate of such correct recognitions rose from 0.72 to 0.87, and overall accuracy increased from 0.57 to 0.65. At the same time, they slightly more often predicted disease control incorrectly.

The next stage of I3LUNG is already underway: the system is being tested in more than 2,000 patients. The published study describes its retrospective phase, the analysis of accumulated clinical data and the review of completed cases together with physicians.