The Dutch Initiative for the Model-Based Selection of Patients for Proton Therapy – A Federated Learning IT Infrastructure

Abstract

<h3>Purpose/Objective(s)</h3> Model-based selection for proton therapy (PT) patients has been implemented since 2018 in the Netherlands. Normal tissue complication probability (NTCP) models translate dose differences in Organs at Risk (OARs) (Δdose) into differences in NTCP (ΔNTCP). With these NTCP models, patients that will benefit most from PT are selected for proton therapy. These models are based on patients treated with modern photon techniques due to a lack of data for proton therapy to base them on, assuming that their performance is similar among patients treated with PT. Our primary objective was to test the accuracy of one of these NTCP models in a proof-of-concept study with the external validation of the grade 3+ dysphagia NTCP model (six months after the treatment). <h3>Materials/Methods</h3> We used data from 320 head and neck (HNC) patients who were treated between 2019-2021 with (chemo)radiotherapy in one of the Dutch PT centers. Semantic Web technologies and the Findable Accessible Interoperable Reusable (FAIR) principles were applied to the variables needed for the computation of the NTCP grade 3+ dysphagia model formula in order to be structured in a machine-readable format for the Personal Health Train (PHT) infrastructure implementation. The PHT enables the secure exchange of statistical algorithms (i.e., "trains") between different hospitals without the exchange of patient data. Implementing the closed testing procedure methodology, the statistical algorithm for the federated NTCP external validation was built in the RStudio software. This algorithm constitutes the statistical "train" that will be distributed among the different PT Dutch centers. <h3>Results</h3> We successfully completed and implemented the statistical analysis-"train" for the external validation procedure, establishing the end-to-end IT-infrastructure needed in our center according to the requirements of the federated learning ProTRAIT (PROton Therapy ReseArch regIsTry) IT-infrastructure. The performance of the original NTCP model for grade 3+ dysphagia (which was adopted as the updated selected model from the closed testing procedure) had reasonable discriminative power in the validation cohort (AUC=0.84, sensitivity=0.69, specificity=0.75). <h3>Conclusion</h3> With this work we externally validated the existing HNC grade 3+ dysphagia NTCP model in our PT center using the Personal Health Train infrastructure (PHT). The next steps will focus on expanding more applications of the different components of the model-based approach such as the model-based clinical evaluation in the PT centers, including NTCP-model development and validation in a privacy-preserving federated manner with data from the other PT centers.