934P - Harnessing artificial intelligence on real-world data to predict recurrence in head and neck squamous cell carcinoma patients: The HNC-TACTIC study

Background

Almost half of patients with early/locally advanced (LA) head and neck squamous cell carcinoma (HNSCC) progress or relapse within 2 years. We applied artificial intelligence (AI) techniques to extract clinical variables from electronic health records (EHR) and to predict treatment failure.

Methods

This is an international, observational, retrospective study using free-text and structured data from HNSCC patients’ EHRs. A total of 203 clinical variables were extracted running EHRead®, a multilingual natural language processing (NLP) engine employing SNOMED-CT, from the EHR’s of HNSCC patients from 6 hospitals from Spain, France, and Colombia (2014-2021). Only patients with incident early/LA HNSCC were included. We utilized machine learning to train and 10-fold-cross validate several algorithms to predict progression, recurrence, or death within 2 years after radical treatment, selecting the model with the best performance metrics.

Results

Among 2,034,994 patients screened, 1,209 individuals were identified with early (44%) and LA (56%) HNSCC: 39.7% oral cavity, 32.8% larynx, 23.1% oropharynx, and 4.5% hypopharynx. Median age was 65 years (69.2% male). Main comorbidities were hypertension (50.1%), diabetes (33.3%). Most frequent symptoms at diagnosis were pain (60.7%), dysphagia (25.6%). A total of 484 (40%) patients experienced progression, recurrence, or death within 2 years. The algorithm with the best performance to predict radical treatment failure employed a random forest model, with an area under the ROC curve of 0.709. The main predictors of this model were: Age, stage, alcohol consumption, involved nodes, pain at diagnosis and surgical treatment.

Conclusions

This first phase of the HNC-TACTIC study shows the ability of NLP to extract accurate clinical data from HNSCC patients’ EHRs, and for machine-learning models to predict risk of treatment failure, demonstrating the potential capabilities of AI-powered real-world studies. The next phase will validate and retrain this model on data from 9 countries.

Clinical trial identification

NCT05117775.

Editorial acknowledgement

Legal entity responsible for the study

Savana S.L.

Funding

Savana S. L.

Disclosure

M. López, E. Castillo: Financial Interests, Institutional, Full or part-time Employment: Savana S. L. L. Cabal Hierro: Non-Financial Interests, Institutional, Full or part-time Employment: Savana S. L. H.T. Study Group: Financial Interests, Institutional, Full or part-time Employment, Judith Marín-Corral, Bianca de León, Ignacio Salcedo, David Casadevall, Victor Fanjul, Sebastian Menke, Almudena Chapa, Adrián Ceja: Savana S. L. ; Non-Financial Interests, Institutional, Affiliate, John de Almeida (UHN Canada), Andreas Dietz (Leipzig, Germany), Robert Ferris (UPMC, USA, Raul Giglio (Hospital Roffo, Argentina), Chris Holsinger (Stanford University School of Medicine, USA), Kate Hutcheson (MD Anderson Cancer Center, USA), Sandro Porceddu (Queensland Health, Australia), Christian Simon (Lausanne University Hospital, Switzerland), Ana López Alfonso (HUIL, Spain), Gloria María Serrano Montero (HUIL, Spain), Laura Rodrigáñez Riesco (H La Paz, Spain), Beatriz Castelo Fernández (H La Paz, Spain), Cristina Urbano (H de Granollers, Spain), Jordi Serra Carreras (H Granollers, Spain), Carlos García (H Alma Matter de Antioquia, Colombia), Yair Monsalve (H Alma Matter de Antioquia, Colombia), Javier Alonso Ortega (H de Móstoles, Spain), Gabriela Morales Medina (H de Móstoles, Spain), Renaud Schiappa (Centre Antoine Lacassagne, France,: Savana S. L. M. Taberna: Financial Interests, Institutional, Full or part-time Employment: Savana S. L. All other authors have declared no conflicts of interest.