Visual acuity after proton beam therapy for uveal melanoma: development and validation of a prediction model

Aim To develop proton beam therapy treatment planning dose constraints for uveal melanoma (UM), more knowledge is needed on the impact of dose on toxicity. Here, we identify risk factors and develop prediction models for lower post-therapeutic visual acuity. Methods LUMC patients treated at Holland PTC between 2020-2023 (n=135) and at PSI between 1987-2019 (n=100) for UM were included. Candidate predictors for severe and moderate to severe vision impairment (Snellen <0.1 and <0.3) were established through literature review and expert interviews and selected through univariable analysis. Multivariable Cox models were fitted on a randomly selected 2/3rd of the combined dataset (n=156). Performance of both models was calculated on the training set and on the held-out validation set (n=79). Results The final models included baseline visual acuity, ciliary body dose and tumour location for both visual acuity cut-offs, supplemented by the optic disk dose and T-stage for visual acuity <0.1 and macula dose for visual acuity <0.3. Internal performance was good (c-indices 0.760 and 0.761 for visual acuity <0.1 and <0.3, respectively). In the held-out validation set, performance was acceptable (c-indices 0.657 and 0.689) and calibration was insufficient for the model predicting visual acuity <0.1, but acceptable for the model predicting visual acuity <0.3. Conclusion We developed a model predicting moderate to severe visual impairment with good internal performance and acceptable discrimination and calibration on a held-out test set. This is an important step towards NTCP modelling and the development of dose constraints for ocular proton therapy.