ESTRO 2024 - Abstract Book

S5110

Physics - Radiomics, functional and biological imaging and outcome prediction

ESTRO 2024

between weeks) (Figure 1 A). None of the methods were particularly affected by the degree of missing data on patient level if more than 20% (13 time points) were available. If less than 20% of time points were available imputation accuracy dropped with 10-60% (Figure 1 B). Specifically, missing early time points significantly decreased effectiveness of all methods with average accuracy for patients with data missing in the first 7 weeks between 27 and 54% compared to late time points between 39 and 63%.

Conclusion:

From the evaluated methods, KNN showed to be the best imputation method for longitudinal toxicity data, which had the highest overall accuracy and was most stable for different time points. All methods performed significantly worse when xerostomia scores during treatment were missing, and less than 20% of toxicity scores were available. In future research, the effect on NTCP model development when excluding imputed data based on beforementioned effects