1363P - An AI-derived tool to pre-screen lung cancer candidates for clinical trials

Background

Patient selection in clinical trials is crucial but challenging, traditionally dependent on clinician-led identification that might overlook potential participants. We suggest an artificial intelligence (AI)-enhanced approach for institutional patient screening to improve trial enrolment.

Methods

This retrospective study analysed advanced lung cancer patients who attended Gustave Roussy between Feb 2021-Dec 2023. We automated the extraction and structuring of 83 variables from unstructured patient notes, including demographic, disease characteristics (histology, PD-L1 status, molecular and metastatic status), comorbidities, treatment history, and life status. Automated Data Entry (ADE) used advanced language model techniques, with prompt engineering and tailored few-shot examples, compared against Manual Data Entry (MDE) for accuracy. The primary goal was to assess ADE's accuracy of capturing data critical for determining trial eligibility. The method was tested to identify candidates for a phase I trial targeting advanced lung cancer patients with SMARCA4 alterations, pre-treated with ≥1 line of systemic therapy, and an ECOG status of 0-1.

Results

Of 1,057 patients screened, the median completeness of data per patient was 96.7%, with a processing time of 3 minutes per patient. Among 137 patients with both ADE and MDE, ADE achieved 96% accuracy. Of the 39 patients with SMARCA4 alterations, ADE identified 35 (90%), compared to 25 (64%) by MDE. ADE found 21 living patients with SMARCA4 alterations, 15 of whom had metastatic disease. It further identified 13 patients treated in advanced settings, including 8 with progression on ≥one line of therapy (potential candidates) and 5 on first-line chemo-immunotherapy without progression (potential waiting list candidates). After integrating the latest ECOG status, ADE accurately pre-screened 7 patients for the main criteria and 5 for the waiting list.

Conclusions

The AI-driven ADE method is feasible for pre-screening of lung cancer patients for precision medicine trials, achieving over 90% accuracy in identifying main eligibility criteria. This scalable approach could ensure more precise patient selection and optimize clinical trial matching.

Clinical trial identification

Editorial acknowledgement

Legal entity responsible for the study

B. Besse, Institut Gustave Roussy, Villejuif, France.

Funding

Has not received any funding.

Disclosure

M. Aldea: Financial Interests, Personal and Institutional, Research Funding: AstraZeneca, Sandoz, Amgen. P. Rolland, A. Poplu, A. Djarallah, L. Chutto, B. Vignal, S. Platano: Financial Interests, Personal, Affiliate: Lifen. F. Le Ouay: Financial Interests, Personal, Ownership Interest: Lifen. All other authors have declared no conflicts of interest.