In brief
- AI-HOPE Lung Cancer is a multicenter registry across 21 European centers building machine learning models to predict progression-free survival, overall survival, and treatment-related toxicity in metastatic non-small cell lung cancer (NSCLC) treated with first-line immunotherapy-based regimens.
- As of the paper’s writing, 920 patients had been recruited, with 621 having baseline imaging scans available for centralized analysis.
- The registry combines a retrospective arm reaching back to 2017 with a prospective arm running until 2027, and stores clinical and imaging data on a dedicated platform, the San Raffaele Ai CEnter (S-RACE).
- Planned models combine structured clinical variables with CT and FDG-PET imaging features through a partially automated imaging analysis workflow, alongside explainable AI tools.
- The paper reports the registry’s design and recruitment to date; it does not report prediction accuracy, model performance, or clinical outcome results.
Study at a glance
- Design: Multicenter real-world registry with a retrospective arm (from 2017) and a prospective arm (running until 2027)
- Centers: 21 across Europe
- Population: Metastatic NSCLC patients treated with first-line immunotherapy-based regimens
- Recruited to date: 920 patients, of whom 621 have baseline imaging scans available for centralized analysis
- Primary objective: Time-to-event models for progression-free survival (PFS) and overall survival (OS)
- Complementary models: Binary classification for early progression, long-term survival, and clinically relevant toxicities
- Data platform: San Raffaele Ai CEnter (S-RACE), a privacy-compliant infrastructure following FAIR data principles
- What was not reported: Model prediction accuracy, performance metrics, or clinical outcome results
920 patients recruited across 21 centers, with imaging for two-thirds
AI-HOPE Lung Cancer is designed to bring together clinical and imaging data from patients with metastatic NSCLC on first-line immunotherapy-based treatment, drawing on a retrospective cohort going back to 2017 and a prospective arm intended to run until 2027. As described in the paper, the registry had recruited 920 patients across 21 European centers by the time of writing, and 621 of them had baseline imaging scans available for centralized analysis, a subset the authors will need for any model that incorporates imaging features.
The paper does not report how far along the prospective recruitment phase is relative to its 2027 endpoint, nor does it give a breakdown of patients by country or center. It describes the registry’s scope and current recruitment rather than any outcome or interim analysis.
A privacy-compliant platform built for machine learning, not just data storage
Clinical and imaging data are harmonized and held within the San Raffaele Ai CEnter (S-RACE) platform, which the authors describe as designed around FAIR principles, meaning the data is intended to be findable, accessible, interoperable and reusable. The stated aim is to minimize the manual workload of preparing this kind of multicenter, multimodal dataset for analysis.
Two kinds of model: time-to-event and binary classification
The registry’s primary objective is to develop time-to-event models for progression-free survival and overall survival. Alongside these, the authors plan complementary binary classification models addressing three more specific clinical questions: whether a patient is likely to progress early, whether a patient is likely to survive long-term, and whether a patient is likely to experience clinically relevant toxicity from treatment.
Where imaging is available, the registry’s approach combines structured clinical variables with features drawn from CT and FDG-PET (fluorodeoxyglucose positron emission tomography) scans, through what the authors call a partially automated imaging analysis workflow on the S-RACE platform. The paper states that explainable AI tools will complement these models, though it does not specify which explainability methods will be used or how their output will be validated.
What the paper does not yet show
This paper describes a registry’s design, infrastructure and recruitment; it is not a report of model results. No prediction accuracy, no validation statistics, and no comparison against existing clinical risk tools are given.
“By combining structured clinical variables and multimodal imaging data, the AI-HOPE Lung Cancer study aims to support refined risk stratification and treatment personalization, ultimately facilitating the responsible integration of AI into routine thoracic oncology practice.”The study’s authors, in ESMO Real World Data and Digital Oncology (2026)
An infrastructure paper ahead of the results
What AI-HOPE offers at this stage is scale and structure: a multicenter, multimodal dataset built specifically for machine learning work in metastatic NSCLC, with more than 900 patients already recruited and a defined pipeline for adding imaging analysis. Whether the models it produces meaningfully improve individualized prediction of survival and toxicity over existing clinical tools is a question this paper sets up rather than answers.
Sources
- ESMO Real World Data and Digital Oncology. AI-HOPE lung cancer: a multicenter real-world registry integrating artificial intelligence for metastatic non-small-cell lung cancer (2026-08-03). doi.org
Featuring Lung Summit faculty
This study was co-authored by Lung Summit faculty Roberto Ferrara and Lizza Hendriks.
This article was produced independently by the Lung Summit editorial team, without industry funding or input.
