Safe Active Learning of Cerebrospinal Fluid Dynamics

This paper presents a safe active learning framework for a clinically relevant class of nonlinear systems with time-varying and uncertain parameters. The framework aims to provide a systematic trade-off between three competing clinical objectives: regulation of physiological variables to a safe zone, learning of patient-specific parameters, and minimization of the medical intervention. To address these challenges, we integrate the covariance propagation of a Kalman filter used for patient parameter estimation into an optimization-based control algorithm and enforce a desired estimation accuracy by introducing a soft constraint on the predicted covariances. We demonstrate the potential of the safe active learning framework for healthcare applications in a case study on cerebrospinal fluid dynamics. Our proposed method improves patient monitoring and shunt therapy for the neurological condition hydrocephalus by doubling the parameter estimation accuracy while requiring less than half the rate of intervention compared to standard approaches.