Neural ODE-based disease forecasting from retinal imaging with temporal consistency

Efficient clinical trial recruitment and personalized treatment depend on the ability to predict future disease progression from medical images. However, often there is a lack of well-defined biomarkers that can predict future disease development and a wide inter-subject variation in disease progression speed. We address these issues in the context of predicting the onset of late dry Age-related Macular Degeneration (dAMD) from retinal OCT scans. To model the CDF of future dAMD onset, we propose jointly training an AMD stage classifier with a Neural-ODE that predicts the future disease trajectory. A temporal ordering is imposed that inversely relates the distance from the decision hyperplane of the classifier to the time-to-conversion. Furthermore, we ensure intra-subject temporal consistency by incorporating pairs of longitudinal scans from the same eye during training. Our method is evaluated on a longitudinal dataset comprising 235 eyes (3,534 OCT scans), including 40 converters. The results demonstrate the efficacy of our approach, achieving an average eye-level AUROC of 0.83 in predicting conversion within the next 6,12,18 and 24 months, outperforming several popular survival analysis methods.