Banca de DEFESA: JANICE RODRIGUES FARIAS

Uma banca de DEFESA de MESTRADO foi cadastrada pelo programa.
STUDENT : JANICE RODRIGUES FARIAS
DATE: 09/01/2026
TIME: 10:00
LOCAL: https://teams.microsoft.com/meet/24686269925766?p=Eei6vVT48VWLpJhD82
TITLE:

Artificial Intelligence for Molecular Triage: Predicting EGFR Mutations from H&E Slides in a Brazilian Real-World Lung Cancer Cohort

KEY WORDS:

Artificial intelligence; deep learning; digital pathology; EGFR mutation; non–small cell lung cancer; predictive modeling; molecular triage; H&E slides

PAGES: 100
BIG AREA: Ciências da Saúde
AREA: Odontologia
SUMMARY:

Background: Artificial intelligence (AI)–based approaches have emerged as promising tools in predicting actionable genomic alterations directly from hematoxylin and eosin (H&E)–stained slides, potentially optimizing molecular testing workflows. This study evaluated the performance of the NSCLC Panel EGFR algorithm in predicting EGFR mutation status from H&E slides of non–small cell lung cancer (NSCLC) in an independent Brazilian real-world cohort within a resource-limited setting. Methods: A total of 214 H&E slides from 203 NSCLC patients (107 EGFR-mutant and 107 wild-type) were retrospectively retrieved from pathology archives. After excluding slides with insufficient tumor tissue or poor quality, 181 were included. All slides were digitized and analyzed using the NSCLC Panel AI model, blinded to molecular and clinical data. Model performance was assessed using receiver operating characteristic (ROC) analysis, Youden’s index, and Bayesian post-test probability estimation. Results: The NSCLC Panel achieved an AUC of 0.831 (95% CI, 0.756–0.887; p < 0.001) for discriminating EGFR-mutant from wild-type cases. At the optimal cutoff defined in this study, the model demonstrated 89.6% sensitivity, 68.2% specificity, a positive predictive value of 48.5%, and a negative predictive value of 95.2%, supporting its use as a prescreening tool. Conclusions: The NSCLC Panel demonstrated robust predictive performance and high negative predictive value for EGFR mutation detection in a diverse, real-world Brazilian cohort, comparable to established international models. These findings support AI-assisted prescreening as a feasible strategy to optimize molecular testing resources and enable earlier targeted therapy, particularly in low- and middle-income healthcare settings.

COMMITTEE MEMBERS:
Presidente - 3437282 - ANDRE FERREIRA LEITE
Externo à Instituição - ANDRÉ MÁRCIO MURAD - UFMG
Externo à Instituição - MARCOS ANTONIO DOS SANTOS - HUB
Externa ao Programa - 1609346 - MYLENE CHRISTINE QUEIROZ DE FARIAS - UnB