6527 Background: Survival probability changes as patients survive longer periods of time after diagnosis, and estimates of prognosis at diagnosis are no longer valid. Conditional survival (CS) accounts for the changing hazard rate over time and is a more accurate estimate of survival probability for these cancer survivors. The specific aim of this project was to build a statistical model and web-based tool to predict conditional survival for an individual head & neck cancer patient based on tumor and patient characteristics. Methods: Using 27,825 patients diagnosed with head & neck cancer between 1988–97 from the Surveillance, Epidemiology, and End Results 17 (SEER) database, we built a multivariate Cox proportional hazards regression prediction model. Patient and tumor characteristics included as covariates were age, sex, race, tumor site, stage, and grade. The primary endpoint was conditional overall survival. The model was validated for discrimination using the concordance index and a calibration plot was constructed. Bootstrapping was used to correct for optimistic bias. We also built a web-browser software tool to allow a user to enter patient information into the model and calculate conditional survival probability. Results: The regression model showed very good calibration and discrimination with a bootstrap-corrected C-index of 0.71. For a 65-yr old white male with a moderately-differentiated tonsil cancer with regional lymph nodes, the model predicted that the 5-yr conditional overall survival would increase from 50% at the time of diagnosis to 63% at 3 years after diagnosis. For a 75-yr old black male with a well-differentiated localized lip cancer, 5-yr conditional survival would improve from 58% at diagnosis to 70% by 3 years from diagnosis. Conclusions: Our regression model can accurately predict conditional survival for head & neck cancer patients based upon specific patient and tumor characteristics. This tool allows the calculation of more specific prognosis predictions for individual cancer patients who have already survived a period of time after diagnosis and treatment. No significant financial relationships to disclose.
Cox's Regression Model for Counting Processes: A Large Sample Study