Myo-ODE: continuous-time trajectory reconstruction and risk prediction of high myopia via neural ordinary differential equations.

The global surge in childhood myopia necessitates robust screening tools for early risk stratification; however, conventional predictive models often struggle with irregular follow-up intervals and fail to capture the continuous nature of refractive development. We propose Myo-ODE, a continuous-time framework based on Neural Ordinary Differential Equations (Neural ODEs) to predict high myopia risk. Unlike traditional discrete machine learning models, Myo-ODE parameterizes the derivative of the refractive state, allowing myopia progression to be represented as a continuous latent dynamic flow. This architecture explicitly accommodates non-uniform screening intervals and supports temporal interpolation and cautious short-term extrapolation within the observed follow-up horizon. Evaluated on a longitudinal dataset (N = 4,973), Myo-ODE achieved the highest F1-score of 0.8000 (95% CI: 0.7741-0.8256) and Recall of 0.7812 (95% CI: 0.7518-0.8103), while maintaining a competitive AUC of 0.9834 (95% CI: 0.9781-0.9887). Furthermore, our framework reconstructs individualized refractive progression trajectories and provides an interpretable estimate of the model-learned progression momentum of SE change. By bridging the gap between discrete clinical observations and continuous trajectory-level modeling, Myo-ODE offers a promising tool for personalized myopia surveillance in real-world screening environments.