Session Item

In the Netherlands, patients are selected for proton therapy (PT) using the model-based approach, using normal tissue complication probability (NTCP) models to translate dose differences in OARs (Δdose) into differences in NTCP (ΔNTCP). Currently used NTCP-models for model-based selection are based on patients treated with modern photon techniques, assuming that these models perform similarly among patients treated with protons. An important component of the model-based approach is a continuous update of NTCP-models in patients treated with protons and photons. To support this continuous update of NTCP-models, a national project was initiated to build an national IT-infrastructure to develop and validate NTCP-models (ProTRAIT-project). The basic IT-infrastructure has recently be completed and can be used for model development and validation.

To test the feasibility of this national IT-infrastructure, we performed a proof-of-concept validation of a NTCP-model predicting grade 3+ dysphagia at six months after treatment. Our goal was to establish a proof of concept method for external validation of NTCP-models that can be implemented and extended by the Dutch proton therapy centres without the exchange of patients data implementing the Personal Health Train infrastructure (PHT)¹.

We used a dataset of 65 head and neck (HNC) patients  that were treated between 2019-2021 with (chemo)radiotherapy in MAASTRO clinic. Using semantic web technologies, the data variables needed for the computation of the NTCP model formula of figure 1 were transformed in a machine-readable and Findable Accessible Interoperable (FAIR) data principles² for the implementation of the PHT. Following the closed testing procedure methodology³ the statistical algorithm for the federated NTCP external validation was built in the Rstudio software and constitutes the statistical “train” that will be exchanged among the different proton therapy centres (figure 2).

After the “FAIRification” of the patients data, we successfully implemented the statistical analysis-“train” needed for the external validation in our dataset. The performance of the NTCP model for grade 3+ dysphagia had reasonable discriminative power in the MAASTRO’s cohort (AUC=0.77) with the model update selected as the suitable model update method (AUC=0.90). Furthermore, we established the end-to-end IT-infrastructure needed in our centre according to the requirements of the federated learning ProTRAIT IT-infrastructure.

This study has delivered a proof of concept federated learning infrastructure for the external validation of NTCP models, in which grade 3+ dysphagia was used as an example. Future work will focus on extending more applications of the different components of the model-based approach, including NTCP-model development and validation, model validation and model-based clinical evaluation.

References

1.       https://pht.health-ri.nl/

2.       PubMed ID: 26978244

PubMed ID: 27891652