Theoretical Identifiability of Blood Glucose from the Electrocardiogram: A Real-Data-Validated Multiscale Forward Model and Cramer-Rao Lower Bound Analysis

Le résumé fourni par la source

Existing ECG-based blood glucose detection research is uniformly inverse in approach: collect data, train a model, report performance. This leaves a blind spot. When performance is poor, one cannot tell whether the algorithm is inadequate or whether the signal simply does not carry enough information. This study takes the opposite approach. We build a forward model from established physiology, compute how large a signal a given glucose change actually leaves on the ECG, compare that against measured real-world noise, and derive a Cramer-Rao lower bound that no algorithm can beat. The model is a six-layer multiscale chain spanning milliseconds to days: (M1) glucose trajectory, taken from real CGM recordings rather than synthesised; (M2) counterregulatory epinephrine release with a sigmoid threshold and HAAF adaptation; (M3) extracellular potassium dynamics, including an insulin-independent term; (M4) ventricular repolarisation using the O'Hara-Rudy 2011 human myocyte model, with an effect-compartment delay, a circadian QTc term, and a direct beta-adrenergic pathway in addition to the potassium route; (M5) autonomic modulation of heart rate via an IPFM model; (M6) ECG synthesis with real recorded noise from the MIT-BIH Noise Stress Test Database rather than Gaussian white noise. The central design feature is M7, a set of validation gates. The model must reproduce specific quantitative observations from the published human literature before any identifiability analysis is permitted. Five of the six gates carry veto power. This is the mechanism by which an in-silico study earns the right to make claims about a real technology, and it is why we describe this work as a real-data-validated in-silico analysis rather than a simulation study. Only after all veto gates pass does M8 proceed: Fisher information and the Cramer-Rao lower bound for glucose in the presence of unknown nuisance parameters, Sobol global sensitivity analysis over twelve input factors, and an overlap analysis quantifying the irreducible error rate arising from the non-monotonic mapping in which both hypoglycaemia and hyperglycaemia prolong QT. Five hypotheses are pre-specified with decision rules fixed in advance. All are declared reportable whether or not they are supported; a negative result is explicitly stated to be publishable. This registration covers Study B of a three-study programme. Study C is registered at https://doi.org/10.17605/OSF.IO/KX9JZ. Study A will be registered separately before its data are accessed.

Ce résumé expose les affirmations des auteurs. BNTIC ne l’interprète pas comme une validation indépendante des résultats.