Prediction of early death among patients (pts) enrolled in phase I trials: Development and validation of a new model based on platelet count and albumin level.

Abstract

2540 Background: Selecting pts with “sufficient life expectancy” for phase I oncology trials remains challenging. The Royal Marsden Model (RMM; Arkenau, JCO 2009) identified high-risk pts as those with ≥2 of: albumin (ALB) <35g/l; LDH >ULN; more than 2 metastatic sites. The RMM was assessed in 1845 pts treated in phase I trials by members of the European New Drug Development Network (ENDDN). The present study aimed to develop an alternative prognostic model using a different methodology and to compare its performance with the RMM. Methods: The primary endpoint was 90-day mortality rate (90-DMR). The new model was developed from the ENDDN database using CHAID, an exploratory non-parametric data analysis method evaluating the relationship between a dependent variable (90-DMR) and possible predictive variables through a decision-tree analysis. The ROC characteristics and calibration of both methods were then validated in the independent EORTC Database (n=341 pts). Results: The CHAID method identified low and high-risk groups with 90-DMR of 9.5% vs 37.5%. High-risk pts had ALB<33g/L or ALB≥33g/L but platelet count ≥400.000/mm3. Applying both models to the validation dataset; the rates of correctly-classified pts were 0.86 [CI-95% 0.82-0.90] vs. 0.67 [CI-95% 0.60-0.74], with the CHAID model and RMM respectively. The CHAID model had higher specificity (0.90 vs. 0.65), but lower sensitivity (0.36 vs. 0.93) than the RMM. Discriminative slopes were similar for the CHAID model and RMM (19.4% and 19.3%, respectively). The negative predictive value (NPV; correct identification of pts surviving 90 days) was similar for the CHAID model and RMM (0.94 [0.91-0.97] and 0.99 [0.94-1.00] respectively). Calibration, assessed by the Brier score, was slightly better with the CHAID model (0.001 and 0.098, respectively). Conclusions: In selecting pts for phase I trials arguably an important criterion is NPV; the CHAID model and RMM provided a similar high level of NPV but the CHAID model gave a better rate of correctly-classified patients. Both models improved the screening process and reduce the attrition rate in phase I trials.