An AI-powered complete blood count panel for high-grade cervical lesion identification: A multicentric retrospective study.

e17535 Background: Early detection and accurate risk stratification of high-grade cervical lesions (HSIL) are crucial for effective treatment and improved patient outcomes, as timely interventions are both cost-effective and successful in preventing progression to cancer. In this retrospective, multicentric study conducted across five laboratories in Brazil, we propose leveraging machine learning (ML) to repurpose complete blood count (CBC) tests as a cost-effective tool to identify women at high risk for HSIL. Methods: We analyzed complete blood count (CBC) tests from 324,291 women aged 25–64 years who underwent Papanicolaou tests or cervical biopsies within six months of their CBC test, collected between January 2004 and February 2024, across five Brazilian laboratories: Fleury, Amais SP, Labs Amais, Novamed, and Lafe. Of these, 768 women (0.24%) were confirmed as cases with biopsies identifying HSIL, while 323,523 were classified as controls based on Papanicolaou tests negative for malignancy. A ridge regression model was trained using data from one laboratory's database, with the remaining five databases used for testing. Results: Statistical analysis revealed that mean corpuscular hemoglobin (MCH), mean corpuscular volume (MCV), corpuscular hemoglobin concentration mean (CHCM), lymphocyte, and lymphocyte-monocyte ratio (LMR) were significantly higher (p<0.05) in women with HSIL, while age, red blood count (RBC) and RDW were significantly lower. Furthermore, using a feature selection methodology based in a decision tree, we incorporated age, MCH, LMR, and RDW in a ridge regression model, achieving an average external validation AUC of 0.70 ± 0.02, sensitivity of 0.74 ± 0.04, specificity of 0.63 ± 0.06, accuracy of 0.63 ± 0.06, balanced accuracy of 0.68 ± 0.04, negative predictive value (NPV) of 1.00 ± 0.00, and positive predictive value (PPV) of 0.05 ± 0.02. This indicates that by selecting 37% of the population for screening, we could identify approximately 74% of women with HSIL. Conclusions: Our AI model, utilizing CBC parameters, demonstrates potential as a pre-screening tool to identify women at elevated risk for HSIL, thereby optimizing the allocation of Papanicolaou or HPV DNA testing resources. This strategy may be particularly advantageous in resource-limited settings, where access to comprehensive screening programs may be constrained. To support its clinical application, external validation across diverse populations is essential to ensure the model's generalizability and effectiveness.