ESTRO 2024 - Abstract Book

S5109

Physics - Radiomics, functional and biological imaging and outcome prediction

ESTRO 2024

QLQ-H&N35). Scores from 16 time points were included: before, weekly during treatment (7 time points), and at 12 weeks, 6, 12, 18, 24, 36, 48 and 60 months after treatment. This data was split into training (70%), validation (15%) and test (15%) sets. Multiple imputation methods were included: the relatively simple mean and mode imputation, the more complex K Nearest Neighbour (KNN) and Multiple Imputation by Chained Equations (MICE). All methods except MICE were fitted on the training set and then applied to validation and test set. For MICE, the full dataset was supplied to the procedure. All evaluations were done on the test set. To assess the imputation performance an available toxicity value was removed and then imputed in the test set. This process was repeated for all time-points, i.e. removing and predicting the toxicity value for all patients per time point. Two metrics were used to compare imputed and true values: accuracy for absolute agreement (range: 0-100%) and normalized Root Mean Squared Error (nRMSE) as an error measurement (higher being worse). Moreover, performance of the methods was additionally evaluated: 1) per time point, 2) the degree of missing data on patient level and 3) influence of certain time points to be available.

Results:

Overall accuracy with KNN and MICE was 66.4% (95%CI: [58.0-74.8]) and 60.9% (95%CI: [51.3-70.5]) respectively, which was better than that obtained with mean and mode imputation in which the overall accuracy was 39.6% (95%CI: [27.1 52.1]) and 44.1% (95%CI: [31.5-56.7]) respectively. Similar results were found for nRMSE, which was 0.012 and 0.015 for KNN and MICE, respectively. And 0.018 and 0.019 for mean and mode, respectively (Table 1).

MICE performed worse for late time points (for week 1-7, average accuracy 68%; after week 7, and average accuracy 49%) whereas KNN imputation was more robust for missing values at all time points (no more than 5% difference

Made with FlippingBook - Online Brochure Maker