Page 185 - Functional impairment and cues for rehabilitation of head and neck cancer patients -

was a consistent predictive factor in other studies (16, 24, 25). In our study, univariable analysis showed a significant association but this association was no longer present in multivariable analysis. An explanation for this might be that in some previous studies dysphagia was assessed by means of validated questionnaires whereas in our study the variable was less reliably obtained from notes in the medical file.
The second consistent predictive factor for tube dependency in literature is advanced nodal stage (16, 22, 25, 26, 38). The influence of this factor could not be estimated in our cohort due to the lack of variation among nodal stages (70% of the patients had stage N2). This, however, will most likely not influence generalizability because the other variables in the model are independent predictive factors.
Previous studies suggest a predictive value of radiotherapy dose [16, 26, 39]. We, however, did not use parameters such as bilateral neck irradiation or constrictor dose as predictors, because we aimed to develop a prediction model based on readily available clinical parameters, easy to use in daily practice. Despite the advantage of this model concept, the absence of such radiotherapy data has to be considered one of the limitations of this study.
A few studies have made an attempt at predicting the risk on prolonged tube feeding dependency (19, 21, 27, 39). In contrast to ours, these included heterogeneous populations with regard to tumor stage and treatment modalities with only a proportion of patients being treated with primary CRT (17-60%), which compromises their generalizability to the CRT population. Also, all studies included patients who received a prophylactic feeding tube, which puts them at risk for selection bias.
The strengths of this study include the large dataset with all patients treated with primary CRT for a broad range of tumor localizations compared to previously published work, which enables the construction of an accurate prediction model. Also, a reactive placement protocol was used for all patients. In case a prophylactic placement strategy would have been applied in a proportion of patients, the risk of selection bias would have been high.
Limitations of the study include its retrospective nature. We do not think that the results are affected by this because the number of missing variables was low and most likely random. A prospective study design is preferred in developing a prediction model. However, considerable amount of time is needed to include sufficient patients prospectively. The current data is therefore the best available and the only way to enable risk estimation of prolonged feeding tube dependency. Another limitation of this study is the lack of consistent criteria used in practice to decide the timing of tube feeding, which contributed to disagreement of observed and predicted probabilities. Moreover, some misclassification of high and low risk patients can also be explained by patients who refused tube feeding.
7
Prediction model for tube feeding during CRT
183