Evaluating imputation methods for longitudinal clinical trajectories: Toxicities in head and neck cancer treatment.

PURPOSE: Longitudinally collected medical data is often afflicted by missingness, making methods mitigating missingness valuable. This study evaluated nine imputation methods for longitudinal data, using toxicity scores measured at multiple timepoints resulting from head and neck cancer treatment with radiotherapy. METHODS AND MATERIALS: For four toxicities (xerostomia, dysphagia, taste alteration, and sticky saliva), toxicity scores at 16 timepoints (1 prior, 7 during, and 8 after treatment) were collected, resulting in 1442 patients with on average ∼11.6 available timepoints. Imputation methods ranged from simple to deep learning-based. Methods were evaluated using a systematic removal strategy enabling comparison of true and imputed values. Patient-level, cohort-level and downstream model evaluation was done, evaluating inclusion criteria for patients, training data requirement and modelling effect respectively. RESULTS: General imputation quality showed minimal differences between methods, except simple methods, which performed worse. For patient-level evaluation, it was found that when patients had <50% of timepoints available, imputation started to worsen. Cohort-level evaluation showed that advanced methods could reach good performances at low patient counts (150 patients) but were more severely affected by low average timepoint availability. Downstream model evaluation showed negligible differences between methods, other than simple methods which resulted in different model coefficients. CONCLUSIONS: It was found that it is best to avoid simple imputation methods. Furthermore, when applying imputation to longitudinal data practitioners should be conscientious about inclusion of certain patients and if their dataset permits the use of more advanced imputation methods.