Skip to content

RRational Validation Summary

Last update: 2026-06-03 · RRational version: 0.9.3

This page documents the scientific validity of the HRV pipeline in RRational. It combines (a) the published basis of NeuroKit2 (the underlying engine), (b) an independent cross-validation against Kubios HRV Scientific, and © guideline conformance (Task Force 1996 / Quigley 2024).

1. Why NeuroKit2 as the engine?

NeuroKit2 (Makowski et al. 2021) is a widely used, peer-reviewed open-source Python library for physiological signal processing. RRational uses it for all core HRV calculations.

Main NeuroKit2 publication

Makowski, D., Pham, T., Lau, Z. J., Brammer, J. C., Lespinasse, F., Pham, H., Schölzel, C., & Chen, S. H. A. (2021). NeuroKit2: A Python toolbox for neurophysiological signal processing. Behavior Research Methods, 53(4), 1689–1696. DOI: 10.3758/s13428-020-01516-y

What the paper documents: - Open source and peer-reviewed, with R-peak detection algorithms benchmarked against open ECG databases (the benchmark results live in the package's online validation documentation rather than in the article itself). - Implements the standard time- and frequency-domain HRV indices used throughout the field.

The HRV index definitions RRational relies on (band boundaries, RMSSD/SDNN/pNN50 formulas) follow the field-standard references directly — Task Force (1996) and Shaffer & Ginsberg (2017) — not the NeuroKit2 paper, which does not itself cite those guidelines.

Shaffer, F., & Ginsberg, J. P. (2017). An overview of heart rate variability metrics and norms. Frontiers in Public Health, 5, 258. DOI: 10.3389/fpubh.2017.00258

Supporting NeuroKit2 / HRV methodology literature

  • Pham, T., Lau, Z. J., Chen, S. H. A., & Makowski, D. (2021). Heart rate variability in psychology: A review of HRV indices and an analysis tutorial. Sensors, 21(12), 3998. DOI: 10.3390/s21123998
  • Frasch, M. G. (2022). Comprehensive HRV estimation pipeline in Python using Neurokit2: Application to sleep physiology. MethodsX, 9, 101782. DOI: 10.1016/j.mex.2022.101782

Bibliometric context: As of 2026 the NeuroKit2 paper has accrued >2,000 citations (Google Scholar), reflecting broad adoption across physiological-signal research. This figure is provided as approximate, time-stamped context, not as a claim from the paper itself.

2. RRational vs Kubios HRV Scientific — independent cross-validation

To check that RRational produces values comparable to an established reference tool, we processed identical RR-interval sequences from 5 participants through both RRational and Kubios HRV Scientific v4.3.0 and compared the resulting metrics.

The underlying recordings contain identifiable participant data and are therefore not redistributed with the repository. The exact pipeline parameters are documented below so the comparison can be reproduced on equivalent recordings.

Test setup

  • Input: The same RR-interval sequences were provided to both tools, with no tool-specific re-segmentation.
  • RRational mode: freq_method="kubios" — the Kubios-aligned pipeline added in v0.9.3.
  • Kubios settings: Defaults as documented in the Kubios batch-export CSV header.

To make the two tools comparable, RRational's Kubios-aligned mode mirrors the Kubios processing chain. RRational's default (freq_method="neurokit") instead follows NeuroKit2 conventions (notably normalized spectral power), which are not directly comparable to Kubios absolute-power output — see the Kubios Compatibility Guide.

Processing step Kubios HRV Scientific v4.3.0 RRational (freq_method="kubios")
Beat correction Automatic threshold (0.3) or none, per export NeuroKit2 Kubios algorithm (Lipponen & Tarvainen 2019)
Detrending Smoothness Priors, λ = 500 Smoothness Priors (Tarvainen et al. 2002), λ = 500
Interpolation Cubic spline, 4 Hz Cubic spline, 4 Hz
PSD estimator Welch FFT, 180 s window, 50% overlap Welch FFT, 180 s window, 50% overlap
Spectral scaling Absolute power (ms²) normalize=False → absolute power (ms²)
Frequency bands VLF 0–0.04, LF 0.04–0.15, HF 0.15–0.40 Hz Identical

Result

Agreement on four cleanly matched segments (rest_pre, B11_SB, B9_SB, B12_U), expressed as the absolute relative difference between the two tools:

Metric Mean |Δ%| Max |Δ%| Rating
MeanNN 0.35% 0.6% excellent
RMSSD 0.83% 1.3% excellent
HF (ms²) 2.7% 4.4% excellent
LF (ms²) 5.4% 8.9% good
LF/HF 5.8% 12.1% good
SDNN see note 30–50% (by design) not directly comparable

SDNN: why the two tools differ by design

The SDNN row above is expected to differ and is not an error — the two tools compute SDNN on different signals:

  • RRational's SDNN (the value RRational reports) is the standard deviation of the raw NN intervals after artifact correction. This is the Task Force (1996) definition and is the internationally cited reference standard.
  • Kubios's SDNN is computed on a detrended, interpolated signal (a proprietary variant). Detrending removes low-frequency variance, which lowers SDNN by roughly 30–50%.

Both quantities are internally valid, but they measure differently defined things and must not be compared directly. For guideline-conformant reporting, use RRational's Task-Force SDNN. Only fall back to a Kubios-style detrended SDNN when the explicit goal is to reproduce Kubios output. Always state which definition was used.

Caveat: pNN50 is unstable by design

When beat correction is active (Kubios "Automatic correction" or RRational nn_correction.method: kubios), pNN50 can differ several-fold between tools even when RMSSD, MeanNN, LF and HF agree closely. From our tests:

  • Kubios "Automatic correction": pNN50 = 16.45% vs RRational 49.86% (~3× difference)
  • Kubios "none": pNN50 = 18.23% vs RRational 17.97% (within ±2%)

This is a fundamental property of the metric, not a bug: pNN50 is a binary threshold counter (only successive differences > 50 ms are counted). Any single corrected beat landing just above or below the 50 ms threshold flips the count.

Rohr et al. (2024) quantified this sensitivity as the relative error per 1 ms of added beat-to-beat noise (standard deviation), normalized to the group mean:

Metric Sensitivity (% error / ms noise SD) Robustness
LF 0.24% very robust
SDNN 0.61% robust
HF 0.71% robust
RMSSD 1.57% moderate
pNN50 2.75% most sensitive

Rohr et al. (2024) conclude that pNN50 "should be used with caution, in particular when the baseline values are expected to be low".

Recommendation: Prefer RMSSD (it carries the same parasympathetic information and is ~4× more robust). If pNN50 is reported, always document the correction pipeline. For Kubios-comparable pNN50 values, set BeatCorrection=none in both tools. See the Kubios Compatibility Guide.

Additional sources on pNN50 / RMSSD behaviour:

  • Berntson, G. G., Lozano, D. L., & Chen, Y.-J. (2005). Filter properties of the root mean square successive difference (RMSSD) for heart rate. Psychophysiology, 42(2), 246–252. DOI: 10.1111/j.1469-8986.2005.00277.x
  • Mietus, J. E., Peng, C.-K., Henry, I., Goldsmith, R. L., & Goldberger, A. L. (2002). The pNNx files: re-examining a widely used heart rate variability measure. Heart, 88(4), 378–380. DOI: 10.1136/heart.88.4.378
  • Alcantara, J. M. A., Plaza-Florido, A., Amaro-Gahete, F. J., et al. (2020). Impact of using different levels of threshold-based artefact correction on the quantification of heart rate variability in three independent human cohorts. Journal of Clinical Medicine, 9(2), 325. DOI: 10.3390/jcm9020325

Interpretation in the context of published tool comparisons

Inter-tool HRV agreement is known to be imperfect in the literature, especially for frequency-domain metrics, because tools differ in PSD implementation, detrending and band handling:

Tool (vs Kubios) Reported agreement
pyHRV (Gomes et al. 2019) 26 of 78 HRV parameters significantly different from Kubios
hrv-analysis (Champseix et al. 2021) "Some small differences observed can be explained by mathematical approximations"
RRational (this cross-validation) LF |Δ| ≤ 9%, HF |Δ| ≤ 4% on matched segments

Conclusion: With frequency-domain differences below 10% across matched segments, RRational's Kubios-aligned mode sits at the favourable end of the inter-tool agreement reported in the open-source HRV literature.

Limitations

This is a proof-of-concept cross-validation on 5 participants, not a formal multi-cohort validation study. Segment selection was limited to cleanly matched recordings, and the comparison was performed against a single reference tool (Kubios v4.3.0). The numbers above should be read as evidence of pipeline equivalence under Kubios-aligned settings, not as population-level validity claims.

3. Guideline conformance

RRational implements recommendations from:

Task Force 1996 (gold standard)

Task Force of the European Society of Cardiology and the North American Society of Pacing and Electrophysiology (1996). Heart rate variability: standards of measurement, physiological interpretation and clinical use. Circulation, 93(5), 1043–1065. DOI: 10.1161/01.CIR.93.5.1043

Implemented: - Frequency bands: VLF 0.003–0.04 Hz, LF 0.04–0.15 Hz, HF 0.15–0.40 Hz - Time-domain: SDNN, RMSSD, pNN50 with standard definitions - Recording length: the Task Force recommends a nominal 5-minute short-term recording (or 24-hour long-term); it does not define a 2-minute time-domain minimum. RRational's own minimum-data thresholds (≥100 beats for time-domain, ≥300 beats for frequency-domain) follow common practice and the artifact-rate guidance of Quigley (2024).

Quigley 2024 (updated recommendations)

Quigley, K. S., Gianaros, P. J., Norman, G. J., Jennings, J. R., Berntson, G. G., & de Geus, E. J. C. (2024). Publication guidelines for human heart rate and heart rate variability studies in psychophysiology — Part 1: Physiological underpinnings and foundations of measurement. Psychophysiology, 61(9), e14604. DOI: 10.1111/psyp.14604

Implemented: - Artifact reporting with a quality-grade system - Exclusion threshold of >10% artifacts - Full pipeline reporting (detection method, correction, segment length)

Lipponen & Tarvainen 2019 (artifact correction)

Lipponen, J. A., & Tarvainen, M. P. (2019). A robust algorithm for heart rate variability time series artefact correction using novel beat classification. Journal of Medical Engineering & Technology, 43(3), 173–181. DOI: 10.1080/03091902.2019.1640306

Implemented: NeuroKit2's intervals_process(method="kubios") uses this algorithm.

Tarvainen et al. 2002 (Smoothness Priors detrending)

Tarvainen, M. P., Ranta-Aho, P. O., & Karjalainen, P. A. (2002). An advanced detrending method with application to HRV analysis. IEEE Transactions on Biomedical Engineering, 49(2), 172–175. DOI: 10.1109/10.979357

Implemented in Kubios-compatible mode (freq_method="kubios"): nk.signal_detrend(method="tarvainen2002", regularization=500).

Berntson et al. 1997 (HRV committee consensus)

Berntson, G. G., Bigger, J. T., Eckberg, D. L., Grossman, P., Kaufmann, P. G., Malik, M., Nagaraja, H. N., Porges, S. W., Saul, J. P., Stone, P. H., & van der Molen, M. W. (1997). Heart rate variability: Origins, methods, and interpretive caveats. Psychophysiology, 34(6), 623–648. DOI: 10.1111/j.1469-8986.1997.tb02140.x

4. One-page summary for students

RRational is scientifically valid because:

  1. Engine: It uses NeuroKit2, a peer-reviewed, widely cited standard library.
  2. Cross-validation: An independent comparison with Kubios HRV Scientific (5 participants, Kubios-aligned settings) shows agreement of |Δ| ≤ 6% for MeanNN/RMSSD/HF and |Δ| ≤ 9% for LF — at the favourable end of published open-source tool comparisons.
  3. Guideline-conformant: All calculations follow Task Force 1996, Quigley 2024, and Lipponen/Tarvainen 2019.
  4. Transparent pipeline: Every .rrational file records the artifact correction, quality grade, and detection method used.
  5. Optional Kubios mode: If a publication requires direct comparison with Kubios values, enable freq_method="kubios".

For the SDNN and pNN50 differences in your own validation, see the SDNN and pNN50 notes above and the Kubios Compatibility Guide.