Your HRV study, from one iPhone.

RR intervals were recorded with a Polar H10 (recording path not stored with this recording; RR resolution in the data: 1 ms) and processed with CARLON HRV Research, version 1.4, build 2026102801 (Rohwedder, 2026). The recording comprises 5544 RR intervals over 2871.2 s (sum of intervals), starting 2026-10-23 17:09:00 Europe/Berlin (UTC+02:00; start time source not recorded). Beats were classified with the algorithm of Lipponen and Tarvainen (2019) (α = 5.2, 91-beat windows, c₁ = 0.13, c₂ = 0.17; implementation verified beat for beat against NeuroKit2 signal_fixpeaks), with both thresholds bounded below at 30 ms so that beat-to-beat changes smaller than 30 ms, typical of exercise, are not classified as artefacts (see the app's validation report). 0.97 % of beats were flagged. Each run of consecutive flagged intervals was replaced by equal intervals of the same total duration, their number given by whole multiples of the local rhythm where every interval of the run is one (missed beats), otherwise by the classification (one fewer per extra beat, otherwise the same, so a premature beat and its compensatory pause are averaged); normal beats were left unchanged and the duration of the recording was preserved, except for single flagged intervals without compensation, which were replaced by the mean of their neighbours. The per-beat classification and correction are included in the complete export. Mean heart rate was computed as 60 000 / mean RR over all original intervals; SDNN, RMSSD, pNN50 and the Poincaré descriptors SD1 and SD2 (Brennan et al., 2001) were computed on the corrected series. Spectral power was estimated over the whole corrected series with Welch's method (symmetric Hann window of 256 samples = 64 s, 50 % overlap, segment mean removed, density scaling) after natural cubic-spline resampling at 4 Hz; band powers were integrated exactly at the band limits (LF 0.04–0.15 Hz, HF 0.15–0.40 Hz; Task Force, 1996). Baevsky's stress index is reported as √SI with SI = AMo / (2 · Mo · MxDMn) (Baevsky, 1984): histogram of the corrected series with 50-ms bins on the absolute grid (…, 600–650, 650–700 ms …; Baevsky and Kubios convention); Mo = centre of the tallest bin (s), AMo = its share of beats (%), MxDMn = range of the series (s); from 10 min of recording the median over consecutive 5-min windows. Short-term fractal scaling (DFA α1) was computed over windows of 4–16 beats on the whole recording without prior detrending (Peng et al., 1995). Time-varying DFA α1 (Figure 8) followed Rogers et al. (2026): 120-s windows recomputed every 5 s on the corrected series after smoothness-priors detrending (Tarvainen et al., 2002; λ = 500, the Kubios default, as the review does not specify λ); windows with 5 % or more corrected beats were omitted, between the 3 % level with small bias and the 6 % level with proportional bias in Rogers et al. (2021) and in line with the exclusion criterion of Schaffarczyk et al. (2022). Segments (Figure 7) and the 60-s epochs of the data export use fixed time grids; a beat belongs to the window in which its interval ends. The RMSSD trace in Figure 1 uses a sliding window of 30 beats as a display aid only. The app's computations were validated against NeuroKit2 (331 of 331 checks passed; CARLON HRV Research validation report). Original RR intervals were retained unaltered for export.

Recording and storage. RR intervals are recorded with CARLON HRV Research (iOS app) from Polar H10 or Movesense Flash chest sensors via Bluetooth, live or in the sensor’s internal memory with later read-out; until read-out, such a recording remains on the sensor. The app processes and stores the data on the study iPhone. The app has no user account, no server of its own and no usage analytics or advertising services. Data leave the device only through an explicit export by the study team (iOS share sheet) to a destination the team chooses, or through the optional study folder: a folder the team selects in the Files app, e.g. institutional storage, to which the app copies each new recording. Recordings are part of the device’s regular iPhone backup.

Pseudonymization. Participants are entered under a study code chosen by the team (e.g. P-01). On export, the participant can be replaced by a stable pseudonym (ANON- followed by eight hexadecimal characters, derived from an internal identifier with SHA-256). The key linking code and person is kept by the study team outside the app. The data are therefore pseudonymized, not anonymized. Free-text notes are exported as entered; the study team decides what they contain.

Integrity and deletion. Original RR intervals are stored as recorded (physiologically impossible values are discarded and counted); corrections are kept separately. Changes are logged in an append-only audit trail following ALCOA+ principles; a SHA-256 hash chain makes later changes to the log detectable. Recordings can be deleted in the app at any time; exported files, study-folder copies and backups must be covered by the study team’s own deletion procedure.

Classification. The app is a research tool, not a medical device, and is not intended for diagnosis or treatment.