Vancomycert: A Certified Neuro-Symbolic Drug Delivery System
Alistair Sirman, Fleur Conway, Jessica Ciupa, Gusts Gustavs Grīnbergs, Ekaterina Komendantskaya, Alessandro Bruni, Michael John Williams, Vaishak Belle, Thai Son Hoang, Michael Rawson
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 maint aining safety for each specific patient.