Purchase this article with an account.
Mathias Gallardo, Marion Munk, Thomas Kurmann, Sandro De Zanet, Agata Mosinska, Mark van Grinsven, Clara I Sanchez, Martin Sebastian Zinkernagel, Sebastian Wolf, Raphael Sznitman; Machine learning to predict anti-VEGF treatment response in a Treat-and-Extend regimen (TER).. Invest. Ophthalmol. Vis. Sci. 2020;61(7):1629.
Download citation file:
© ARVO (1962-2015); The Authors (2016-present)
To explore the capabilities of a machine learning approach to predict treatment response of patients with wAMD, DME and RVO treated according to a TER.
We performed a retrospective study including 532 eyes (464 patients) with wAMD, 206 eyes (200 patients) with RVO and 262 eyes (194 patients) with DME, treated with anti-VEGF according to a predefined TER during 2014-2018. Eyes were grouped for each disease into good, mediocre and poor responders, using the average treatment interval (good≥8 weeks, poor ≤5 weeks, mediocre: remaining eyes). We trained two Random Forest models to infer the probability of a patient being a good or poor long-term responder. Both models use as inputs morphological information automatically extracted from the OCT B-scans of the three first visits, the three first prescribed treatment intervals and patient-specific information. The training and evaluation of the models were performed using a 10-fold cross-validation on the retrospective cohort while ensuring that, in every split, no patient was present in both training (wAMD:~478, RVO:~185 and DME:~236) and test sets (wAMD: ~54, RVO:~21 and DME:~26).
We observed the capability of predicting, after the third visit, good and poor treatment response for patients with wAMD, RVO and DME with similar performance. In the cohort of wAMD patients, we identified 137 good-, 78 poor- and 317 mediocre responders. For both models and over the 10 folds, a mean AuC of 0.68 (best model over the folds: 0.77) and 0.75 (0.85) was achieved respectively. Models on RVO and DME showed similar results with a mean AuC of 0.7 (0.97) and 0.82 (0.95) and respectively a mean AuC of 0.74 (1) and 0.63 (0.79).
We showed that machine learning classifiers are able to predict the long-term treatment response. This approach may be useful to tailor individual treatment plans for patients.
This is a 2020 ARVO Annual Meeting abstract.
This PDF is available to Subscribers Only