Purchase this article with an account.
Stylianos Serghiou, Bertil E Damato, Armin R Afshar; Use of machine learning for prediction of ocular conservation and visual outcomes after proton beam radiotherapy for choroidal melanoma. Invest. Ophthalmol. Vis. Sci. 2019;60(9):962.
Download citation file:
© ARVO (1962-2015); The Authors (2016-present)
While local control rates for proton beam radiotherapy (PBRT) treatment of choroidal malignant melanoma (CMM) are well known, visual acuity outcomes can vary, and are currently roughly predicted by ophthalmologists based on tumor size and location. The purpose of this study is to apply machine learning methods on a large CMM patient cohort, to predict visual acuity and ocular salvage after PBRT.
This was a prospective cohort study of 1022 adult CMM patients. Our database of 169 features included patient demographics, medical history, ophthalmic history, tumor dimensions, tumor-node-metastasis (TNM) stage, histology, genetics, radiation dosimetry and visual acuity (VA). We predicted final VA and enucleation using cross-validation (CV) to tune and fit the following machine learning methods: Elastic Net, Bayesian Generative Learning, Gradient Tree Boosting and Deep Neural Networks. We finally evaluated our models and a stacked ensemble of our models using cross-validated R2 for VA and cross-validated area under the curve (CV-AUC) for enucleation.
Our cohort consisted of otherwise predominantly healthy adults (834/1022, 82%) with a median age of 59 years (Interquartile Range (IQR), 48-68 years) followed for a median of 8 years (IQR, 3-13 years). Most patients presented with TNM Stage I disease (500/1022, 49%) and VA between 6/5-6/12 (700/1022, 69%). Final vision varied substantially from 6/5-6/12 (384/1022, 38%) to hand-motions or less (114/1022, 11%); 76 (7%) received secondary enucleation. VA was best predicted by Gradient Tree Boosting with a CV-R2 of 28% (95% Confidence Interval (CI), 19-38%) with dosimetry values and CV-R2 of 25% (95% CI, 18-32%) without. The top three most important variables were tumor thickness, radiation received by the macula and radiation received by the globe volume. Enucleation was best predicted by the Elastic Net with a CV-AUC of 80% (95% CI, 72-87%); using dosimetry values did not improve predictive ability. The top three most important variables were tumor thickness, smallest tumor diameter and TNM stage.
Our approach identified models with high predictive ability for enucleation and moderate predictive ability for final visual acuity. Variables related to physical extent of tumor and radiation received appeared most important in predicting enucleation and VA, respectively.
This abstract was presented at the 2019 ARVO Annual Meeting, held in Vancouver, Canada, April 28 - May 2, 2019.
This PDF is available to Subscribers Only