Prognostication System for Squamous Cell Carcinoma Using Retrieval Augmented Generation–Enabled Large Language Model
Menée à partir d'une revue systématique de la littérature et de données portant sur 2 379 carcinomes épidermoïdes cutanés diagnostiqués entre 1996 et 2023 (âge médian : 73 ans ; 61,5 % d'hommes), cette étude évalue, par rapport aux systèmes BWH et AJCC8, la performance d'un grand modèle de langage avec génération augmentée par récupération pour établir un pronostic
Cutaneous squamous cell carcinoma (CSCC) is the second most common cancer in the US. While most tumors are cured with surgery, approximately 4% will develop nodal metastases (NM) and 1% to 2% of patients will die of their disease (disease-specific death [DSD]). Given the annual incidence, CSCC risk stratification is essential to identifying tumors at risk for poor outcomes.There are several methods to risk stratify cancer, including staging systems (ie, TNM systems [tumor, node, metastasis]), prognostic models, and molecular studies. For CSCC, there are 2 widely used staging systems: American Joint Committee on Cancer Staging Manual, eighth edition (AJCC8; TNM staging) and Brigham and Women’s Hospital (BWH; T staging only) T staging system. Performance of both T staging systems has been assessed by several cohorts in the US and Europe, and BWH has been shown to be superior to AJCC8 primary because only 1 risk factor upstages to AJCC8 high stage, so twice as many tumors are considered high stage compared to BWH. However, up to 30% of metastases occur in low-stage tumors by both systems. Another limitation of the staging systems is that tumors in a given stage are assigned equal risk by definition; however, tumor risks vary for several reasons. First, risk factors do not have equivalent associated risks but are treated as much in current systems. Second, arbitrary cut points are used for some risk factors, such as diameter and large-caliber nerve invasion but are continuous in practice. Finally, BWH T2b includes CSCCs with 2 or 3 risk factors, but outcomes risks vary by risk factor number.
JAMA Dermatology , éditorial, 2025