pipette
ENEnglish

Recording sampling rate as a negative-control exposure for detecting measurement-driven bias in heart-rate-variability-outcome associations: a quantitative bias analysis across ICU waveform databases

H. Ye, Q. Zhang, S. Chen, P. Li, J. Wan, Y. Hao

Preprint

En palabras de los autores

Importance. Heart-rate variability (HRV) is studied extensively as a prognostic marker in critical illness, yet the electrocardiographic sampling rate at which HRV is computed systematically distorts the resulting values. Whether the associations between HRV and clinical outcomes reported in the literature are partly a consequence of this distortion has not been examined directly. Objective. To evaluate whether between-database differences in native ECG sampling rate can be used as a negative-control exposure to detect measurement-driven bias in HRV-outcome associations, and to quantify the sensitivity of a published association to an empirically measured distortion coefficient using quantitative bias analysis (QBA). Design. A methods investigation in three parts. Part I sets out the causal structure and states why a strict instrumental-variable (IV) design is inappropriate here. Part II derives a bias-correction factor under a multiplicative measurement-error model and propagates the empirical confidence interval of the distortion coefficient as a sensitivity band. Part III applies the procedure to an observed HRV-mortality association from a single ICU waveform database, and examines the separate bias arising from threshold-based categorisation. Setting and data sources. Two critical-care waveform databases with different native sampling rates (MIMIC-IV Waveform, approximately 250 Hz; MIMIC-III Waveform Matched p04, 125 Hz) and an independent 500-Hz reference cohort (PTB-XL) for mechanism validation. The outcome association was estimated in 1,445 ICU stays from MIMIC-III p04, of whom 141 died in hospital (9.8%). Exposures. Native ECG sampling rate as a negative-control exposure; SDNN, RMSSD and heart rate as the measured exposures whose associations with mortality are evaluated. Main outcomes and measures. In-hospital mortality. Distortion coefficients were obtained by end-to-end reprocessing of 7,293 original waveform segments with a fixed QRS detector. Corrected odds ratios were reported as QBA sensitivity bands rather than point-identified effects. Misclassification was quantified as the proportion of recordings below a 20 ms HRV threshold reclassified above it by coarse sampling. Results. End-to-end reprocessing gave an RMSSD distortion coefficient of 1.1575 (95% CI 1.0835-1.2492; 2,276 segments in 188 patients) for 250->62.5 Hz and 1.0588 (95% CI 1.0294-1.1033; 4,988 segments) for 125->62.5 Hz; high-frequency power was more affected than low-frequency power (1.2033 vs 1.0716). Under the multiplicative model, an observed crude RMSSD-mortality odds ratio of 1.802 corresponded to a corrected interval of 1.8335-1.915 across the coefficient's 95% CI. The correction was small relative to confounding: the same association fell to 1.116 after age adjustment. Coarse sampling reclassified 33.70% of low-HRV recordings in MIMIC-IV and 20.72% in MIMIC-III, a larger potential source of bias than the continuous-scale correction. The cross-database negative-control gradient was completed after linking both waveform cohorts to hospital outcomes: RMSSD inflation was reproducible (median x1.101 at 125->62.5 Hz in MIMIC-III and x1.270 at 250->62.5 Hz in MIMIC-IV), but the age-adjusted HRV-mortality risk ratio did not increase at lower rates (MIMIC-III 1.70->1.21; MIMIC-IV 1.31->1.29->1.20; all 95% CIs crossed 1). The falsification test was therefore negative in these cohorts, and we report it as such rather than as evidence that measurement bias is absent. Conclusions and Relevance. Sampling rate is a detectable and correctable source of bias in HRV-outcome research. Recording the sampling rate of each signal, stratifying effects by baseline variability, and reporting a QBA sensitivity band before dichotomising HRV would make published associations more interpretable. The object of inference here is the measurement-error method, not a new clinical biomarker; the proposed reporting steps require no additional data collection.

Resultado principalLimitación que admiten los autores

Apareció: miércoles, 18 de noviembre. medRxiv. Preprint, todavía sin revisión por pares.

DOI: 10.64898/2026.11.15.26368154