Nature Medicine: AI study says it beat usual tests at guessing who benefits from lung-cancer immunotherapy

Nature Medicine published a study saying an AI tool beat the usual bedside tests at guessing who with a common type of lung cancer will benefit from immunotherapy — drugs that help the immune system attack the tumor.

Doctors already use crude tests to decide who gets these immune-system drugs. A Nature Medicine paper now puts a named, multi-hospital AI study on the record — with the authors’ own scores, including an honest drop when the model left its home hospitals.

The paper, open access under CC BY 4.0, is titled “Clinical usability of an explainable AI decision support tool and evaluation of multimodal models in NSCLC.” NSCLC is non-small cell lung cancer, the most common kind. The performance numbers below are the authors’. This desk did not rerun the models.

The paper is from an international consortium including Fondazione IRCCS Istituto Nazionale dei Tumori di Milano and the University of Chicago thoracic oncology program. Journal metadata lists Arsela Prelaj as corresponding author.

Authors state I3LUNG (trial NCT05537922) is currently the largest international real-world multimodal AI-based study in this setting, enrolling 2,396 patients. That is their claim in Nature Medicine — not this desk’s ranking of every prior cohort.

The model used more than one kind of data: clinical and blood (CB) records, CT scans, digital pathology (DP) slides, and genomics. The authors tried two ways of combining those streams, called machine-learning early fusion (MLEF) and deep-learning intermediate fusion (DLIF).

Using only clinical and blood data, they report an AUC — a 0-to-1 ranking score, where 0.5 is a coin flip — up to 0.77 on a held-out TEST set. That is useful, not perfect. On outside hospitals (EXVAL, external validation) the score fell to 0.55–0.72. They cite population differences for the drop.

On that TEST set, they say the AI significantly surpassed the usual markers: PD-L1 (a protein already used to guess immunotherapy response), ECOG performance status (how well the patient functions day to day), neutrophil-to-lymphocyte ratio (NLR, an inflammation blood ratio), lactate dehydrogenase (LDH, a blood enzyme), and the Lung Immune Prognostic Index (LIPI).

Lung expert and nonexpert physicians improved their guesses when given the explainable-AI clinical-and-blood tool, per the paper.

Adding CT and pathology via MLEF looked better in some analyses. The authors say that extra benefit remains uncertain and did not show up on TEST or EXVAL.

A prospective validation — watching new patients going forward — is underway in more than 2,000 people. This is not approved treatment and not FDA clearance.

CONFIRMED here: the Nature Medicine primary. Distinct from the already-filed deepmind-alphagenome-atlas and openai-codex-chatgpt-antimicrobials — different papers, different tasks.