Control of Medical Digital Twins with Artificial Neural Networks

Lucas Böttcher, Luis L. Fonseca, Reinhard C. Laubenbacher

incompletemedium confidence

Category: Not specified
Journal tier: Strong Field
Processed: Sep 28, 2025, 12:56 AM
arXiv Links: Abstract ↗PDF ↗

Audit review

The paper clearly defines the ANN control law u(bk, ck; θ), the steady-state loss J1, and the metabolic objective J2 and describes training via a straight-through estimator, but it does not provide formal proofs of the properties asked in (a)–(e). The candidate solution supplies those missing derivations: (a) integer-valued controls, (b) gradient expressions for J1 under the differentiable surrogate, (c) clamping θ to [0,1]^2 and existence of a minimizer (given an explicit saturation assumption), and (d)–(e) ratio calculus and gradient conditions for J2, including time-varying controls via a differentiable metamodel. These arguments are correct under the stated assumptions and consistent with the paper’s setup. Minor caveats: (c) requires explicit saturation (not stated in the paper), and gradients are with respect to the straight-through surrogate rather than the original discontinuous mapping.

Referee report (LaTeX)

\textbf{Recommendation:} minor revisions

\textbf{Journal Tier:} strong field

\textbf{Justification:}

This work offers a clear, practically useful framework for controlling ABMs with ANNs and neural ODEs and demonstrates it on two paradigmatic biomedical settings. The approach is timely and leverages automatic differentiation well. The paper would benefit from minor clarifications (sign conventions for controls, explicit saturation mechanics) and brief mathematical appendices to formalize the gradients and properties used in training. These adjustments would improve clarity without altering the substantive contributions.