AI Model Training and Optimisation for Inflammacheck® Early Lung Cancer Detection Device

Lung cancer is the leading cause of cancer-related death in the UK, with over 35,000 deaths annually. Outcomes remain poor, largely because over 70% of cases are diagnosed at stage III or IV, when curative treatment is often not possible. In contrast, five-year survival exceeds 60% for stage I disease. While low-dose CT screening is available for high-risk patients, many people—including those from minority ethnic groups, never-smokers, and those under age thresholds—fall outside screening eligibility. This results in delayed diagnoses, increased mortality, and persistent inequality. There is a clear need for a non-invasive, accessible diagnostic tool to aid earlier detection of lung cancer in both screened and unscreened populations. Inflammacheck® is a CE-marked handheld breath test device that collects hydrogen peroxide and other breath physiology markers using tidal breathing. It has been evaluated in the VICTORY study, which recruited 316 participants from six diagnostic groups. These included confirmed lung cancer (primarily mid-stage), pneumonia, asthma/COPD, bronchiectasis, ILD, and healthy volunteers. Preliminary analysis using advanced machine learning showed good diagnostic accuracy in distinguishing malignant from non-malignant respiratory disease. However, the VICTORY dataset was limited to a single site and lacked representation of early-stage, screen-detected cancers. A separate ExPeL study conducted in Manchester, also using Inflammacheck®, included early-stage cancers from screening pathways. Combined, these studies highlight the potential of AI-enhanced breath testing but indicate the need for a more representative dataset and model refinement. We now propose a 12-month PPIE-supported extension to VICTORY, adding 140 new participants across Portsmouth and Manchester, bringing the final dataset to 456 individuals. Recruitment will prioritise under-represented groups including early-stage, screen-detected cancers, and patients with a high suspicion of cancer but negative biopsy results. Additional participants will also be recruited across comparator groups to preserve balance. Samples will be collected using standardised protocols and uploaded with structured clinical metadata to a secure analysis environment. The project is structured across five work packages. WP0 covers protocol amendment, governance, and system setup. WP1 (Months 1–6) will deliver prospective recruitment and harmonisation of the new dataset with existing data. WP2 (Months 3–9) will support machine learning and deep neural network development, using signal transformation and standardised pipelines. WP3 (Months 7–11) focuses on model validation, including subgroup performance by age, sex, ethnicity, smoking history, and site. WP4 (Months 1–12) ensures effective project management and regulatory preparation. The model will be evaluated using five-fold cross-validation and a held-out test set. Metrics include AUROC, sensitivity, specificity, PPV, and NPV. A final locked model will be generated for validation, then prepared for UKCA and NICE Early Value Assessment. Outputs will include clinical manuscripts, stakeholder reports, and lay materials co-developed with patient and public contributors. This project supports NHS England’s ambition to diagnose 75% of cancers early and reduce inequalities in access to diagnosis. If successful, it will establish the readiness of breath-based diagnostics to support lung cancer triage in routine NHS practice.