Development and External Validation of Integrated Machine Learning-Based Prognostic Model in Oropharyngeal Head and...

Importance: Patient with head and neck cancer of the oropharynx (HNC-OROP) undergo curative-intent definitive or post-operative radiation therapy. The systemic inflammation response index (SIRI) has independent prognostic capacity in HNC-OROP. We hypothesized that the use of SIRI may produce a parsimonious model of HNC-OROP outcomes. Objective: We aimed to investigate the prognostic utility of systemic inflammatory response index (SIRI) in oropharyngeal head and neck cancer patients who underwent radiation therapy. Design, Setting, and Participants: Random survival forest (RSF) machine learning was used to model survival in 568 oropharyngeal cancer patients in this retrospective cohort study. SIRI was calculated via pre-treatment bloodwork. Model validation was performed in an external cohort of 421 oropharyngeal cancer patients. Exposures: Exposure was curative-intent definitive or post-operative radiation therapy for head and neck cancer of the oropharynx (HNC-OROP). Results: This is a retrospective study with 568 and 421 patients in the Roswell Park and external Ohio State University cohorts. We evaluated full and reduced RSF models and a robust decision tree model. The C-index of the models was 0.758 (RSF full), 0.725 (RSF reduced), and 0.702 (decision tree). The incorporation of SIRI (with performance status and smoking history) into a machine learning model identified three risk-groups that significantly stratified overall survival (p < 0.0001). These findings were validated in the external validation cohort (p = 0.0019). Progression-free survival was also significantly different for the three groups in the validation cohort (p = 0.0025). Conclusions and Relevance: An integrated machine learning model using SIRI, performance status, and smoking history was successfully developed and externally validated in oropharyngeal head and neck cancer patients.