Real-World Eligibility and Clinical Utility of ECG-Based Nocturnal Hypoglycaemia Detection: A TriNetX Population Analysis Coupled with Decision-Analytic Simulation

Le résumé fourni par la source

ECG-based non-invasive detection of hypoglycaemia has been proposed as a low-cost alternative to continuous glucose monitoring (CGM). Published models report encouraging discrimination, but two questions determine whether such a device could work in practice, and neither has been addressed: (1) what fraction of the real-world diabetes population is actually eligible, once patients whose physiology or comorbidity would defeat the detection mechanism are excluded; and (2) at realistic nocturnal hypoglycaemia prevalence, what positive predictive value (PPV) and false-alarm burden would follow. This is a two-stage study. Stage 1 is a cross-sectional real-world population analysis in the TriNetX federated network, estimating the Eligible Fraction (EF) under three levels of strictness using a pre-specified, locked list of 14 exclusion criteria in four categories: (A) conditions that directly abolish the detection mechanism (beta-blockers, diabetic autonomic neuropathy, hypoglycaemia unawareness); (B) conditions that contaminate ECG interpretation (atrial fibrillation/flutter, ischaemic heart disease, heart failure, conduction disorders, cardiomyopathy, pacemaker/ICD); (C) conditions affecting QT or electrolytes (QT-prolonging drugs, CKD stage 3+, electrolyte disturbance, non-potassium-sparing diuretics); and (D) thyroid dysfunction. Stage 2 is a decision-analytic Monte Carlo simulation (10,000 iterations) propagating uncertainty in sensitivity, specificity and time-based hypoglycaemia prevalence into PPV, false alarms per night, the specificity required to achieve at most one false alarm per night, k-of-n debouncing under temporally correlated errors, and decision-curve net benefit. Primary outcomes: Stage 1, EF-strict with Wilson 95% CI; Stage 2, median PPV with 95% interval. Five hypotheses are pre-specified. H-C1: the upper bound of the 95% CI for EF-strict is below 0.50. H-C2: the lower bound of the 95% CI for beta-blocker use exceeds 0.25. H-C3: the posterior probability that PPV is below 0.40 exceeds 0.95. H-C4: the specificity required for at most one false alarm per night exceeds the highest specificity reported in the literature under any validation design. H-C5: with error autocorrelation of 0.4 or above, debouncing retains less than 60 percent of the PPV improvement predicted under an independence assumption. A critical distinction is pre-specified: all predictive-value and false-alarm calculations use TIME-BASED prevalence (the proportion of monitoring time spent below the hypoglycaemia threshold, derived from CGM literature), never PATIENT-level prevalence (the proportion of patients with a recorded hypoglycaemia event, derived from TriNetX). The two differ by roughly an order of magnitude and conflating them invalidates every performance estimate; patient-level prevalence is reported separately. Priors are derived by mechanically specified formulas frozen before execution, and all deviations are logged. Negative results are pre-declared as publishable, and EF is pre-declared to be an upper-bound estimate that will not be corrected upward.

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