Joeky Senders

96 Chapter 5 Predictive analysis The discriminatory performance on the hold-out test set as measured by the C-index set ranged between 0.66-0.70 and between 0.67-0.70 across all models for predicting overall survival and one-year survival status, respectively (Table 2). Among the time- to-event models, the integrated C-index ranged between 0.68-0.70 for predicting subject-level Kaplan-Meier survival curves. The AFT model based on a log-logistic distribution demonstrated the highest discriminatory performance for computing personalized survival curves. Compared to all continuous and binary models, the AFT model demonstrated similar or better discrimination for predicting overall survival and one-year survival probability, respectively. Model calibration varied significantly across all models (Supplementary Figure S2). The traditional CPHR model systematically underestimated survival in the 0.5-0.75 one-year survival probability range, whereas the AFT model showed better calibration, particularly in this clinically relevant interval (Figure 2). 0 25 50 75 100 0 25 50 75 100 Predicted Probability (%) Observed Event Rate (%) Model AFT model CPHR model FIGURE 2. Calibration plot demonstrating a systematic underestimation of survival by the Cox proportional hazards regression model in the 0.5 to 0.75 one-year survival probability range and a well-calibrated accelerated failure time model. Abbreviations: AFT=accelerated failure time; CPHR=Cox proportional hazards regression.