The Translational Gap in AI for Oropharyngeal Squamous Cell Carcinoma: A TRIPOD+AI Scoping Review of Methodological Barriers to Treatment Deintensification.
Authors
Affiliations (6)
Affiliations (6)
- Section of Radiology, Department of Biomedicine, Neuroscience and Advanced Diagnostics (BiND), University of Palermo, Via del Vespro 129, Palermo 90127, Italy (G.S., G.C., F.B., L.L.G.). Electronic address: [email protected].
- Section of Otorhinolaryngology, Department of Biomedicine, Neuroscience and Advanced Diagnostics (BiND), University of Palermo, Palermo, Italy (R.G.).
- Section of Radiology, Department of Biomedicine, Neuroscience and Advanced Diagnostics (BiND), University of Palermo, Via del Vespro 129, Palermo 90127, Italy (G.S., G.C., F.B., L.L.G.).
- Section of Otorhinolaryngology, Department of Precision Medicine in Medical, Surgical and Critical Care (Me.Pre.C.C.), University of Palermo, Palermo, Italy (F.S.).
- Ri.MED Foundation, Palermo, Italy (A.C.).
- Department of Radiology, Montefiore-Einstein Medical Center, Bronx, USA (N.L.).
Abstract
To map artificial intelligence (AI) and radiomics applications in computed tomography (CT), magnetic resonance imaging (MRI), and fluorodeoxyglucose positron emission tomography/CT (FDG-PET/CT) for oropharyngeal squamous cell carcinoma (OPSCC) in the context of human papillomavirus (HPV) status and treatment deintensification, evaluate reporting quality using TRIPOD+AI (Transparent Reporting of a multivariable prediction model for Individual Prognosis Or Diagnosis + Artificial Intelligence), and identify barriers to clinical translation. Following PRISMA-ScR (Preferred Reporting Items for Systematic reviews and Meta-Analyses extension for Scoping Reviews), we searched PubMed/MEDLINE and Scopus (January 2015-December 2025) for studies applying AI/radiomics to CT, MRI, and FDG-PET/CT in histologically confirmed OPSCC, focusing on HPV prediction and treatment deintensification. TRIPOD+AI (27 items, 0-2 scoring) assessed reporting quality. Sixty-one studies were included; 34 addressed HPV prediction, 29 survival or prognosis, and 3 treatment response, with several studies addressing more than one objective. CT was the most frequent modality (49.2%), followed by MRI (29.5%) and FDG-PET/CT (21.3%, all 18F-FDG), with manual segmentation (85.2%) and handcrafted radiomics with machine learning (57.4%) being the most common. HPV models reported areas under the curve (AUCs) of 0.65-0.95, but external validation remained limited (22/61, 36.1%), often with performance decline. TRIPOD+AI revealed critical gaps: missing data handling (fully reported 3.3%, absent 54.1%), calibration (6.6% fully, 91.8% absent), interpretable risk groups (27.9% fully, 68.9% absent), data availability (19.7%), and code availability (11.5%). Most studies (78.7%) restricted analysis to the primary tumor; nodal disease was incorporated in 11/61 (18.0%) and was the sole target in 2/61 (3.3%). CT-, MRI-, and FDG-PET/CT-based AI/radiomics models for OPSCC show promising discrimination, but persistent reporting gaps in calibration, transparency, and risk stratification, plus limited external validation and insufficient nodal coverage, preclude safe clinical implementation. Findings apply specifically to HPV-oriented objectives across CT, MRI, and FDG-PET/CT. Standardized protocols, rigorous validation, and structured reporting are needed before these tools can guide treatment deintensification.