AI-Driven Detection of Lung Cancer: Results of a Simulated Retrospective Deployment across Northeast London

Abstract

Background Detecting lung cancer early remains a significant challenge. Despite screening, the majority of lung cancer patients still present symptomatically with advanced stage disease, with a significant minority presenting via Emergency Departments (ED), with associated poorer outcomes. AI application to electronic healthcare records (EHR) could potentially improve early diagnosis. Methods An AI model trained on retrospective EHRs and socioeconomic data of Northeast London patients was deployed in a retrospective simulation as if it had run monthly from 1/8/2017-1/8/2023. During this period, the model generated monthly risk predictions selecting the 100 highest-risk patients each month selected for simulated pseudo-radiology referral (reflecting real-world capacity) with a maximum of one pseudo-referral per 12-month period. Efficacy was assessed against eventual cancer diagnoses within a 2-year period of the pseudo-referral date, the assumption being that any cancer diagnosed after more than two years would not have been present at the time of pseudo-referral. Stage at ultimate diagnosis, route to diagnosis and time fromAI pseudo-referral to diagnosis were recorded. Results 80,634 patients were included, of whom 7,110 were pseudoreferred with 79 ultimately diagnosed with cancer within 2 years of the AI pseudo-referral. Of pseudo-referred patients with cancer, 58% were ultimately diagnosed at stage III/IV, with 65% diagnosed via A&E/Consultant Upgrade pathways and only 31% diagnosed via GP referral. Mean time from pseudo-referral date to ultimate real-life diagnosis was 460 days (95% CI: [366, 554]) for early-stage and 490 days (95% CI: [427, 553]) for late-stage cancers. Conclusions This AI model shows promise to detect lung cancer many months earlier than standard care. If proven through a prospective trial, earlier detection could shift stage at diagnosis, reduce emergency presentations and improve patient outcomes, lowering healthcare costs. Disclosure This work was funded by AstraZeneca; of whom BB and AC are employees. CT was an employee and shareholder at Clinithink at the time of this research. Copyright © 2026 Elsevier B.V.