Skip to contents

Changelog

Source: NEWS.md

mrpheus 0.1.4.9000 (development)

🚀 Performance fixes in detect_qrs()

detect_qrs() was hanging indefinitely on real recordings (unbounded on a 468,700-sample / ~15.7 min recording at ~500 Hz). Two separate O(n²) bottlenecks were found and fixed:

  • Peak selection (.find_peaks_min_dist()): was checking every candidate peak against every other accepted peak on each iteration — fine on toy data, unusable once candidate counts reach the hundreds of thousands. Replaced with a binary search (findInterval()) against a pre-allocated, sorted buffer of accepted peak locations. Note this is O(n_candidates log k
    • k²) where k is the final peak count (bounded by physiological heart rate, not by recording length), rather than a universal O(n log n) — fine here, but worth remembering if this helper is ever reused somewhere the output peak count could scale with input size.
  • Moving-average integration (Stage 4): used stats::convolve() for a plain boxcar filter. R’s FFT-based convolution can degrade badly depending on how the padded length factors for stats::nextn(), and this was the actual bottleneck once the peak-selection fix was in place. Replaced with a cumulative-sum boxcar implementation, mathematically equivalent to the FFT-based full convolution (verified numerically identical, differences on the order of 1e-15) and unconditionally O(n).

Combined effect: detect_qrs() on the 468,700-sample recording above now completes in 0.97 seconds (1855 R-peaks detected, mean HR 118 bpm).

✨ New function: compute_hrv_freq()

  • compute_hrv_freq() — frequency-domain HRV (HF power) locked to each infant’s own respiratory rate, rather than the fixed adult-derived HF band (0.15–0.4 Hz) that doesn’t apply to neonatal breathing rates (~40–70 breaths/min at rest). Detects the dominant respiratory frequency directly from the physlog’s resp channel, resamples the RR-interval series via cubic spline, and integrates Welch PSD power in a band centered on that detected frequency. Verified against synthetic data: exact frequency recovery and correct amplitude-scaling of integrated power.
  • Fills the gap left by compute_hrv_sleep(), which remains a stub expecting staged PSG input (wrong shape for physlog data) with its RR-to-HRV step never wired in — compute_hrv_freq() is a separate, physlog-shaped path rather than a completion of that stub.

🧪 Tests

  • test-find-peaks-min-dist.R: .find_peaks_min_dist() against a brute-force reference on random and periodic synthetic signals, the minimum-distance constraint, tallest-peak-wins tie-breaking (including the all-tied flat-signal edge case), and a timing regression guard on a large, realistically-smoothed candidate set.
  • test-boxcar-integration.R: the cumulative-sum boxcar formula against stats::convolve(type = "open") across several window sizes and signal lengths, plus an end-to-end detect_qrs() timing smoke test on a synthetic two-minute ECG-like signal.
  • test-compute_hrv_freq.R: synthetic QRS/respiration generators injecting a known respiratory frequency via RSA-modulated RR intervals — frequency recovery at two different frequencies, HF power scaling with RSA amplitude, band-bounds correctness, the too-few-beats NA path, and implausible-RR-interval filtering.

All fixtures are fully synthetic — no participant-derived data.

mrpheus 0.1.4 (2026-07)

🚀 Rcpp hot-path optimisation (staging pipeline + event detection)

Complete C++ hot-path port across the staging and event detection pipelines, replacing the remaining R-level bottlenecks. zoo removed from Imports.

Staging pipeline — per-epoch functions (src/staging_features.cpp):

  • nzc_cpp: zero-crossing count — single O(N) pass replacing sum(diff(sign(x)) != 0)
  • petrosian_fd_cpp: local extrema count + log formula — replaces diff + logical sum
  • hjorth_cpp: mobility and complexity — single O(N) pass accumulating all three variances (x, diff(x), diff(diff(x))) without materialising intermediate vectors
  • stat_features_cpp: std, IQR (type-7 quantile), skewness, excess kurtosis — replaces sd(), IQR() (which sorts), and two mean() calls per epoch
  • rowmedian_cpp: row-wise median of the Welch periodogram matrix — eliminates ~750 000 R median() dispatch calls per recording (251 freq bins x ~3 000 epochs)
  • roll_right_mean_cpp: right-aligned partial rolling mean (k=4), replacing zoo::rollapply for the _p2min_norm normalisation step (49 calls/recording)
  • robust_scale_cpp: (x - median) / (q95 - q5) with type-7 quantile and NA-awareness, replacing median() + quantile() for 49 normalisation calls

Event detection (src/event_detection.cpp, new file):

  • roll_rms_cpp: centered rolling RMS replacing zoo::rollapply in compute_spindles() — processes the full concatenated epoch signal per channel
  • detect_so_candidates_cpp: zero-crossing scan + duration/amplitude filtering for slow oscillation detection, replacing the R which() + lapply loop in compute_slow_oscillations()

🔧 Other changes

  • Fixed a latent bug where the pure-R vapply definition of .roll_triang_mean in the normalisation section was shadowing roll_triang_mean_cpp on every load_all(). Now a single Rcpp-backed definition exists.
  • Removed zoo from Imports — no remaining zoo:: calls anywhere in the package.

🧪 Tests

All 371 tests pass. 0 errors | 0 warnings | 0 notes.

mrpheus 0.1.3 (2026-07)

🩺 PSG preprocessing pipeline

Full preprocessing pipeline from EDF to analysis-ready epochs, matching the companion Python/MNE notebook workflow.

  • preprocess_psg() — continuous-signal preprocessing pipeline: channel renaming, DC removal, powerline notch filter (auto-detected 50/60 Hz + harmonics), and channel-type-specific bandpass filtering (EEG/EOG/EMG/ECG). Operates on the full continuous signal before re-epoching to avoid filter discontinuities at epoch boundaries.
  • remove_dc(), detect_powerline(), notch_filter(), bandpass_filter() — exported signal-level functions; each operates on a plain numeric vector independently of the PSG pipeline.