KEY POINTS
- This retrospective methodological analysis used prospectively collected toxicity data from 1,442 patients receiving primary radiotherapy for head and neck squamous cell carcinoma. Xerostomia, dysphagia, taste alteration and sticky saliva were assessed at 16 time points, spanning pretreatment baseline, seven treatment weeks and follow-up through 60 months.
- Each patient had an average of approximately 11.6 available assessments out of 16, with time-point response rates ranging from 25% to 91%. Investigators compared nine imputation methods, including mean/mode replacement, linear interpolation, knowledge-based interpolation, K-nearest neighbors, MICE, Bayesian ridge regression, a neural network and an assumption-based trajectory method.
- Bayesian ridge regression achieved the highest binary classification accuracy across all four toxicities: 87.4% for xerostomia, 90.0% for dysphagia, 88.9% for taste alteration and 85.5% for sticky saliva. Corresponding normalized RMSE values were 0.137, 0.115, 0.145 and 0.149, respectively.
- Simple mean/mode imputation performed substantially worse, with representative binary accuracies of 68.2% for xerostomia, 71.7% for dysphagia, 74.1% for taste alteration and 70.1% for sticky saliva. By contrast, differences between more advanced approaches were generally small, with no consistent advantage from neural-network complexity.
- Imputation reliability deteriorated when individual patients had fewer than approximately 50% of assessment time points available. Training-data experiments suggested that some advanced methods could perform adequately with approximately 150 patients, provided longitudinal data completeness remained relatively high.
- Despite differences in imputation accuracy, downstream six-month NTCP model performance was largely unchanged across non-simple methods. For xerostomia, complete-case analysis achieved AUC 0.73 and Brier score 0.20, compared with AUC 0.72–0.74 and Brier score 0.20–0.21 after imputation; simple methods were more likely to alter model coefficients.
- The study evaluated imputation by artificially removing observed values, approximating missing completely at random conditions. Real clinical missingness can depend on symptoms, illness severity, treatment burden or death, so the results may not fully generalize to missing-at-random or missing-not-at-random patterns; the authors also did not incorporate additional clinical covariates into the imputation models.
CLINICAL TAKEAWAY
For longitudinal head and neck toxicity datasets, relatively straightforward methods such as Bayesian ridge regression may achieve performance comparable to or better than more complex machine-learning approaches, while mean/mode imputation should generally be avoided. Method selection must account for patient-level data completeness and missingness mechanisms, and better imputation accuracy should not be assumed to produce better NTCP prediction.