ACUMEN: Active Cross-Entropy Method with Uncertainty-driven Neural ODEs for Data-Efficient System Identification in Healthcare

Amirhossein Afkhami Ardekani, Mahmoud Zeydabadinezhad, Babak Mahmoudi

Keywords: Active learning, Neural ODEs, Uncertainty quantification, System identification, Healthcare time series, Deep brain stimulation, Model predictive control, Cross-entropy method, Epistemic uncertainty, Digital twins, Precision medicine, Data-efficient learning, Ensemble methods

TL;DR: We present ACUMEN, an uncertainty-driven active learning framework combining Neural ODEs with CEM-MPC to achieve 24% better sample efficiency in healthcare system identification by exploring high-uncertainty regions.

Abstract:

Building personalized, data-driven models of patient response to therapies with large parameter spaces,
such as Deep Brain Stimulation, is a major challenge in precision medicine. We present ACUMEN, a
data-efficient framework for digital twin–inspired modeling that couples Neural Ordinary Differential
Equations (Neural ODEs) with uncertainty-driven planning via Cross-Entropy Method Model Predictive
Control (CEM-MPC). An ensemble of Neural ODEs captures physiological dynamics, with ensemble disagreement
quantifying epistemic uncertainty. CEM-MPC selects exploratory interventions under smoothness constraints,
aided by optimistic state progression and adaptive scaling across state dimensions.
In the Reinforcement Learning for Deep Brain Stimulation (RL-DBS) synthetic environment, ACUMEN reduces test error by up to 24.2\% over passive data collection, while producing tighter uncertainty estimates. These results highlight ACUMEN’s potential to lower sample complexity and enable safer, more personalized system identification in data-limited healthcare-inspired settings.

Submission Number: 86