Carotid UHF‑RFID Heartbeat Detection Master Dataset (Subjects 1–15)

A passive UHF RFID sticker tag (860–960 MHz, anti-metal type) was affixed to the skin overlying the right carotid artery and its S11 return loss measured with a NanoVNA-H V3.6 vector network analyzer through a circularly polarised 902–928 MHz reader antenna. Arterial wall displacement beneath the tag perturbs antenna–tissue impedance matching periodically, so the cardiac rhythm appears directly in the tag's reflection coefficient without electrodes, conductive gel, or any powered element on the body. S11 was sampled at 20 Hz for 30 s per trial (600 samples), and heart rate was recorded once at trial onset by pulse oximetry (SparkFun MAX30101 PPG sensor). The dataset comprises 175 trials over 35 subject–position combinations: five trials in each of sitting, standing, and lying for Subjects 1–5, and five trials in sitting and standing for Subjects 6–15. Subject-level records are limited to age, height, sex, self-reported fitness (1–5), and brief health notes; no identifying information is released. Data collection spans three acquisition strata, which are released as a covariate and should not be pooled without accounting for their structure. Subjects 1–5 were recorded on the original front end and show narrow, high-amplitude spikes. An improved noise-cancellation stage was then introduced: Subjects 6–10 were recorded after it with the S11 minimum tracked over 860–930 MHz, and Subjects 11–15 with tracking restricted to 890–920 MHz. Beat-resolution reliability diverges sharply at the Subject 11 boundary, coincident with the band restriction, though the acquisition records do not establish it as the cause. Acquisition cohort is perfectly confounded with subject identity, as no subject was recorded under more than one configuration, and the strata are not demographically balanced — Subjects 11–15 are all female, aged 45–50. Two subjects are of particular interest. Subject 5 shows the highest rates in the cohort (107–121 BPM by position). Subject 13 is pacemaker-dependent and supplies a near-fixed-rate control rhythm, making it the most useful single validation case in the set. This version includes visual summaries and statistical graphics—such as Bland–Altman plots and subject‑specific analytical metrics—to illustrate multiple aspects of the dataset. The original wide‑format dataset remains available under its own versioned DOI. Suitable for: time-domain, frequency-domain, or learned heart-rate extraction from S11 oscillations, benchmarked against the recorded onset BPM; postural effects on RF-based cardiac sensing; inter-subject generalisation; and study of respiratory contamination and estimator failure modes, since a respiratory component near 0.29 Hz (17.2 ± 1.2 breaths/min) is present throughout. Note that the reference is a single trial-onset reading rather than a beat-by-beat label, so agreement statistics computed against it are bounded below by its own uncertainty and cannot be read as beat-level accuracy. Data files Carotid RFID Heartbeat Detection Master Dataset: Subjects 1–15 (Processed Signal, Metadata, and Beat‑Resolution Annotations).csv — the dataset proper. 104,964 rows, one per sample. Column Description subject Subject number, 1–15 acquisition_cohort One of the three strata described above position Sitting, Standing, or Lying trial Trial number within subject and position, 1–5 sample_index Sample number within trial, from 0 time_s Seconds from trial start, 0–29.95 at 0.05 s spacing s11_return_loss_db S11 return-loss magnitude in positive dB; larger values indicate a better antenna–tissue match bpm_reference Trial-onset pulse-oximetry heart rate, repeated down the trial for ease of joining subject_metadata.csv — one row per subject: age, height, sex, self-reported fitness (1–5), health notes, acquisition cohort. per_trial_metrics.csv — one row per trial (175 rows): per-window S11 statistics (mean, min, max, peak-to-peak range, standard deviation), both recovered rate estimates, interbeat-interval coefficient of variation, whether an interior cardiac-band spectral peak was present, and the dominant sub-cardiac (respiratory) peak. cohort_agreement.csv — Bland–Altman agreement statistics by estimator and cohort: trial count, mean absolute percentage error, bias, 95% limits of agreement, and estimate-to-reference ratio. The two derived files are the output of an independent reimplementation of the estimators described in the accompanying data descriptor, not of the original analysis code. They agree with the published results on nearly every point, but disagree on one: for Subjects 6–10 this implementation locates an interior peak in the 0.7–2.5 Hz cardiac band in 50 of 50 trials, scattering around the reference at 0.93 ± 0.21, where the descriptor reports no such peak in any trial and a collapse onto the 0.7 Hz band edge. Resolve this against the original code before treating either file as authoritative. Figures Each figure is a vector PDF. Colour denotes acquisition cohort throughout: Subjects 1–5 in red, 6–10 in gold, 11–15 in blue. Sitting_all_subjects_waveforms.pdf, Standing_all_subjects_waveforms.pdf, Lying_all_subjects_waveforms.pdf — one file per body position. Each opens with a grid page showing every subject recorded in that position, five overlaid trials per panel, annotated with mean reference heart rate. Subsequent pages give one subject each: the five overlaid waveforms at full width above, and below them a per-trial comparison of the pulse-oximetry reference against both recovered estimates. The three cohort morphologies are visible directly in the grid pages — narrow spikes for Subjects 1–5, smooth low-frequency oscillation with no resolved beats for Subjects 6–10, and discrete beat-scale features for Subjects 11–15. The lying file covers only Subjects 1–5, the sole subjects recorded in that position; postural and acquisition effects are therefore fully confounded for lying. fig04_estimate_vs_reference.pdf — recovered rate against the pulse-oximetry reference for all trials, by cohort, for (a) the threshold-based peak counter and (b) the spectral estimator. Dashed line marks perfect agreement, dotted the half-rate locus, and in (b) a dash-dotted line marks the 0.7 Hz search-band edge. Subjects 6–10 cluster tightly below the half-rate locus under the peak counter rather than scattering across it. fig05_bland_altman.pdf — Bland–Altman analysis of estimate against reference for both estimators, with bias and 95% limits of agreement drawn per cohort. Limits are computed within cohort rather than pooled, since strata differing in front end are not comparable. fig06_cohort_mean_spectra.pdf — (a) cohort-averaged autocorrelation spectra on a log scale, with the 0.7–2.5 Hz cardiac search band shaded and the respiratory peak near 0.29 Hz marked; (b) per-trial ratio of cardiac-band peak power to respiratory peak power, in dB. The cardiac component sits roughly 26 dB above respiration for Subjects 1–5, about 9 dB below it for Subjects 11–15, and about 24 dB below for Subjects 6–10. fig07_bpm_by_position.pdf — mean trial-onset heart rate by subject and position. Standing exceeds sitting for every subject with multi-position coverage, and sitting exceeds lying for three of the four with released lying waveforms. fig08_mean_s11_by_position.pdf — mean S11 by subject and position, with error bars giving between-trial standard deviation. Between-trial spread is small relative to between-subject spread, indicating good repeatability, but mean S11 does not order consistently by posture, so position should be treated as a covariate rather than assumed to shift the baseline in a fixed direction. fig09_signal_quality_by_cohort.pdf — distributions of per-trial peak-to-peak range, within-trial standard deviation, and interbeat-interval coefficient of variation, by cohort. Interbeat CV is elevated for Subjects 6–10 (0.31–0.60) but does not cleanly separate the cohorts on its own, as Subjects 1–5 span 0.00–0.42. Licence Creative Commons Attribution 4.0 International (CC BY 4.0).