Vancomycert: A Certified Neuro-Symbolic Drug Delivery System (Case Study)

2026-06-17

AI Summary

This work addresses the challenge of ensuring both adaptability and provable safety over an infinite time horizon for neural network controllers in clinical settings. The authors develop a clinically interpretable, simplified physiological model and train a neural network via supervised learning to emulate vancomycin dosing strategies. For the first time, they formally verify the controller’s infinite-time safety using the Rocq language and the Vehicle interactive theorem prover. The end-to-end proof guarantees that the controller’s outputs never exceed drug toxicity thresholds, thereby ensuring that blood concentration remains within the therapeutic window at all time steps. This approach effectively mitigates nephrotoxicity risk while accommodating personalized and diverse dosing regimens.

Abstract

Neural network controllers for autonomous decision-making are well-established in cyber-physical systems, yet their deployment in safety-critical healthcare settings remains largely unverified. This paper presents a methodology and case study for the formal verification of a neural network controller for antibiotic dosing, motivated by the challenge of systems that must be simultaneously adaptive and provably safe across unbounded time horizons. We construct a simplified yet clinically-interpretable model that tracks drug concentration, body temperature, and white blood cell count. Vancomycin is selected as a representative antibiotic, widely prescribed for severe infections yet carrying a narrow therapeutic window, where supratherapeutic concentrations risk nephrotoxicity and subtherapeutic dosing risks treatment failure. A supervised neural network controller is trained on synthetic clinician-style dosing data. We establish formal verification of input-output safety properties, specifically verifying a property of a neural network that implies an infinite-horizon proof that automated dosing never exceeds the supratherapeutic boundary. This system property is proven in Rocq using the Vehicle interactive theorem prover back-end to integrate the different proof systems. The end result is a verification pipeline that allows for a wide variety of treatment approaches whilst maintaining safety for each specific patient.

Problem

• neural network controller
• formal verification
• antibiotic dosing
• safety-critical systems
• therapeutic window

Innovation

• neural-symbolic verification
• formal safety guarantees
• infinite-horizon verification
• antibiotic dosing control
• interactive theorem proving