Quantifying the Robustness of Accelerometer-Derived Gait Features for Step Counting Across Sensor Locations and Sampling Frequencies

Review PDF
Denario-0
2508.00023-R1 📅 14 Apr 2026 🔍 Reviewed by Skepthical GitHub

Official Review

Official Review by Skepthical 14 Apr 2026
Overall: 4.8/10
Soundness
4
Novelty
5
Significance
5
Clarity
6
Evidence Quality
4
The paper tackles a practical and timely question with a simple, interpretable pipeline, and the core empirical trends are coherent; however, rigor is undermined by major methodological gaps and errors. The mathematical audit flags a critical ENMO formula inconsistency, the numerical audit finds an unresolved mismatch in step-event totals, and the evaluation pools heavily overlapping windows without participant-wise analysis, uncertainty quantification, or clear direction handling for AUC—limiting claims of robustness and generalization. Data provenance, annotation/synchronization details, feature and DSP specifications, and an ethics statement are insufficient, and the failed demographic merge raises additional data-integrity concerns. Despite these issues, the observed patterns (variability features strongest; 25 Hz often sufficient) are plausible and useful, yielding moderate significance but requiring substantial revision for technical soundness and evidence completeness.
  • Paper Summary: This manuscript evaluates how robust nine simple accelerometer-derived time- and frequency-domain features are for discriminating “Step” vs “Non-Step” $2~\mathrm{s}$ windows ($50\%$ overlap) across two sensor locations (hip, wrist) and two sampling frequencies ($100~\mathrm{Hz}$, $25~\mathrm{Hz}$) (Sec. 2.1–2.4). Triaxial acceleration is converted to ENMO, windowed, labeled as “Step” if at least one annotated step event falls inside the window, and reduced to scalar features including magnitude/variability summaries (mean, SD/variance, IQR), peak-based metrics, and spectral summaries (dominant frequency, spectral energy in $0.5\text{--}3~\mathrm{Hz}$, spectral entropy) (Sec. 2.2–2.3). Each feature is assessed as a univariate classifier via ROC-AUC for each condition (Sec. 2.4; Fig. 1; Table 1). The main empirical pattern is that variability/magnitude features (especially SD/variance, IQR, and band-limited spectral energy) yield high AUCs, remain comparatively strong at $25~\mathrm{Hz}$, and generally perform better at the hip than at the wrist, while simple peak-based features degrade markedly at the wrist and some spectral features behave inconsistently (Sec. 3.2–3.4, 3.6). The paper is clearly organized and the pipeline is simple and interpretable, but key methodological details are currently under-specified (data provenance, step annotation/synchronization, preprocessing and exact feature definitions). In addition, the evaluation pools hundreds of thousands of overlapping windows without participant-wise evaluation or uncertainty quantification (Sec. 2.4.1, 3.1–3.4), which limits claims about robustness and generalization. A planned demographic subgroup analysis failed due to a metadata merge error (Sec. 2.4.3, 3.5), and an internal inconsistency in reported step-event totals (Sec. 2.1 vs Sec. 3.1) needs correction.
Strengths:
Addresses a practically important “bigger picture” question for wearable deployment: which simple features remain reliable under sensor placement (hip vs wrist) and reduced sampling ($100~\mathrm{Hz}$ vs $25~\mathrm{Hz}$) (Sec. 1).
Simultaneous hip and wrist recordings within the same $39$ participants enable clean within-subject comparisons across locations and sampling rates, reducing confounding (Sec. 2.1).
Simple, transparent pipeline (ENMO $\rightarrow$ fixed windows $\rightarrow$ predefined feature set $\rightarrow$ AUC) supports interpretability and makes trends easy to communicate (Sec. 2.2–2.4).
Results yield actionable guidance: variability/magnitude features and $0.5\text{--}3~\mathrm{Hz}$ spectral energy are consistently strong, whereas naive peak counting is unreliable at the wrist (Sec. 3.2–3.4, 3.6; Fig. 1).
Limitations are at least acknowledged (notably the failed demographic merge), providing a clear set of next steps (Sec. 3.5, 4).
Major Issues (8):
  • Data provenance, protocol, and ground-truth step annotation/synchronization are under-specified, limiting reproducibility and external validity (Sec. 2.1, Sec. 2.2.2, Sec. 3.1). It is unclear what study the $39$ participants come from, what activities/environments were included (e.g., treadmill vs free-living; walking only vs mixed tasks), how step events were obtained (manual video, foot sensors, device algorithm), annotation resolution/quality control, and how hip and wrist streams (and $100$ vs $25~\mathrm{Hz}$ streams) were time-aligned and drift-corrected. These details are critical because labels are derived from step timestamps.
    Recommendation: Expand Sec. 2.1 and Sec. 2.2.2 to document: (i) dataset/study origin (new vs reused; cite source if reused), setting and activity protocol, recording durations, inclusion/exclusion; (ii) step annotation method, definition of a “step event,” temporal resolution, QC/validation; (iii) synchronization/alignment procedure across sensors and sampling rates (timestamping, drift handling). Explicitly describe how ambiguous segments (turns, transitions, stairs, shuffling, arm-only motion) were handled, and discuss likely label noise implications in Sec. 3.6/4.
  • Unit of analysis and dependence: AUC is computed once per feature/condition on pooled, overlapping windows across all participants (Sec. 2.4.1, Sec. 3.1–3.4). With $50\%$ overlap and strong within-subject temporal correlation, treating $\sim$hundreds of thousands of windows as independent can inflate effective sample size, obscure between-participant heterogeneity, and overstate “robustness”/generalization to unseen individuals.
    Recommendation: Revise Sec. 2.4.1 to adopt a participant-wise evaluation: report per-participant AUCs (or leave-one-subject-out / participant-wise $k$-fold CV) and summarize distributions (mean$\pm$SD or median[IQR]) for each feature and condition in Sec. 3.2–3.4. If you keep a pooled AUC, also report participant-level results as the primary robustness evidence and use pooled results only as a descriptive complement.
  • No uncertainty quantification or formal comparisons for AUC differences (Sec. 2.4.1–2.4.2, Sec. 3.2–3.4, Sec. 3.6). Claims about “minimal degradation,” “robustness,” or location/sampling effects rely on point estimates without confidence intervals/tests, making it unclear which differences are statistically/practically meaningful.
    Recommendation: Add uncertainty estimates using a clustering-respecting method: e.g., bootstrap resampling by participant (not by window) to obtain $95\%$ CIs for AUCs and key contrasts (hip vs wrist; $100$ vs $25$; top features vs others). If using participant-level AUCs, perform paired comparisons across participants (with multiple-comparison control where appropriate). Update Sec. 3.6 and Sec. 4 to ground robustness statements in these intervals.
  • Task/label definition creates a mismatch with “step counting” claims: windows are labeled “Step” if $\geq1$ step occurs in a $2$ s window (Sec. 2.2.2), which conflates step presence with step density and “walking-like” movement. Many features (SD/IQR/spectral energy) will scale with the number of steps and intensity within the window; windows with $1$ step vs $4$ steps are treated identically, and mixed-activity windows are possible. This limits how directly conclusions translate to event-level step detection/counting (Sec. 1, Sec. 3.6, Sec. 4).
    Recommendation: Clarify throughout (Abstract, Sec. 1, Sec. 3.6, Sec. 4) that the evaluated task is window-level discrimination of step-containing windows (a proxy for walking/movement bouts), not end-to-end step counting. In Sec. 2.2.2 and Sec. 3.1, report the distribution of steps-per-positive-window and discuss mixed-window prevalence and boundary effects. If feasible, add a small sensitivity analysis: alternative labeling rules (e.g., $\geq2$ steps; or majority-of-samples walking), and/or a regression to predict step count per window, to test whether the main feature rankings and “$25~\mathrm{Hz}$ sufficiency” claims hold.
  • Feature definitions and signal-processing details are incomplete, and some are internally inconsistent across sampling rates (Sec. 2.2.1–2.3.2). In particular: (i) ENMO is written with gravity subtraction inside the square root (dimensionally inconsistent) (Intro p.2; Sec. 2.2.1 p.2–3); (ii) it is unclear whether ENMO is truncated at $0$ (common ENMO definition); (iii) peak detection is under-specified (min distance, edge handling, filtering/smoothing) and is likely sampling-rate sensitive; (iv) FFT features lack details (demeaning/DC handling, window function, scaling/normalization, bin selection for $0.5\text{--}3~\mathrm{Hz}$ at $25$ vs $100~\mathrm{Hz}$); (v) spectral energy wording conflicts (“power spectrum” vs “squared magnitudes,” potentially implying $|\mathrm{FFT}|^4$) (Sec. 2.3.2).
    Recommendation: Fix and standardize definitions in Sec. 2.2.1–2.3.2: (a) correct ENMO with explicit parentheses and units, e.g., $\mathrm{ENMO} = \max(0, \|\mathbf{a}\| - 1g)$ if truncation is intended; use the same expression everywhere; (b) specify whether $25~\mathrm{Hz}$ is native or downsampled, and if downsampled, detail anti-alias filtering/decimation (Sec. 2.2.2); (c) fully specify peak detection algorithm and parameters (library/function, prominence, min peak distance in seconds, preprocessing, boundary treatment) (Sec. 2.3.1); (d) fully specify FFT pipeline (demean/detrend, DC exclusion, window type, zero padding, PSD normalization) and give explicit formulas for dominant frequency, spectral energy (as a band power/integral), and spectral entropy (including probability normalization and log base) (Sec. 2.3.2). Ensure frequency-band implementation is comparable across $25/100~\mathrm{Hz}$ (bin mapping).
  • The demographic subgroup analysis failed due to a metadata merge/parsing error, but the failure mode and data-integrity implications are not sufficiently diagnosed (Sec. 2.4.3, Sec. 3.1, Sec. 3.5). Without clear verification, readers cannot be sure only demographics were impacted (and not participant IDs, file-to-subject mapping, or label alignment).
    Recommendation: In Sec. 2.4.3 and Sec. 3.5, provide a concrete diagnosis (which keys failed, how IDs differed, what rows were missing/duplicated). Explicitly document integrity checks that confirm the main feature–label dataset is correct (participant-level counts, file hashes/checksums, sanity checks on time ranges). If any participants/sessions were excluded, list them. If feasible, fix the merge and include at least a minimal subgroup summary; otherwise, strengthen generalization caveats in Sec. 4.
  • Internal inconsistency in reported step-event totals: $48,721$ (Methods Sec. 2.1) vs $62,904$ (Results Sec. 3.1). This undermines trust in dataset accounting and downstream results.
    Recommendation: Reconcile these numbers by explicitly stating scopes/filters for each count (e.g., before/after exclusions, per-condition vs combined, annotation revisions). Correct the manuscript so the same quantity is reported consistently, and ensure any derived percentages in Sec. 3.1 match the corrected totals.
  • Conclusions about robustness and the sufficiency of $25~\mathrm{Hz}$ are broader/stronger than warranted given the described cohort ($39$ healthy adults), unclear activity context, missing subgroup analyses, and the current pooled-window evaluation (Sec. 3.6, Sec. 4).
    Recommendation: In Sec. 3.6 and Sec. 4, restrict claims to the study conditions once fully specified (protocol, population, annotation method). Rephrase statements implying broad deployment across “diverse wearable applications” to note that validation is still needed in older adults, clinical/pathological gait, and multi-day free-living settings. Tie any “$25~\mathrm{Hz}$ is sufficient” claim to specific features and to participant-level uncertainty estimates (after addressing the evaluation issues above).
Minor Issues (8):
  • Windowing/labeling choices are not justified or stress-tested: $2~\mathrm{s}$ windows with $50\%$ overlap (Sec. 2.2.2) may smooth short bouts and create mixed-content windows, affecting AUC and feature rankings; overlap further increases dependence.
    Recommendation: In Sec. 2.2.2, justify $2~\mathrm{s}$ relative to plausible cadence ranges and intended deployment. If feasible, add a brief sensitivity analysis (e.g., $1~\mathrm{s}$ and/or $4~\mathrm{s}$ windows; different overlaps) and summarize whether the main rankings and hip/wrist and $100/25$ conclusions remain stable (Sec. 3.6).
  • AUC direction handling is unclear for features expected to be lower during walking (e.g., spectral entropy as described) (Sec. 2.3.2, Sec. 2.4.1). The text’s interpretation of AUC as $P(\mathrm{feature}_\mathrm{step} > \mathrm{feature}_\mathrm{nonstep})$ can conflict with “lower is more step-like.”
    Recommendation: In Sec. 2.4.1, state explicitly how direction is handled (e.g., compute ROC with the appropriate inequality per feature; or report $\max(\mathrm{AUC}, 1{-}\mathrm{AUC})$; or negate features like entropy). Ensure plots/tables use a consistent convention.
  • Comparability across sampling rates is not explicitly validated for sampling-sensitive features (peak_count, spectral energy) (Sec. 2.3, Sec. 3.4). Without careful normalization, observed “$25~\mathrm{Hz}$ sufficiency” may reflect definition/implementation artifacts rather than true robustness.
    Recommendation: Add a short paragraph in Sec. 2.3.2 and/or Sec. 3.4 explaining why each frequency-domain measure is sampling-rate invariant under your implementation (or how it is normalized to approximate a band-power integral). If $25~\mathrm{Hz}$ is downsampled from $100~\mathrm{Hz}$, consider reporting a controlled downsampling experiment to isolate sampling effects from device/hardware differences.
  • Figure 1 (and Table 1) currently emphasize point estimates and are hard to use for “robustness deltas” (Sec. 3.2–3.4). The grouped bars are dense, and there are no uncertainty indicators.
    Recommendation: Revise Fig. 1 to show uncertainty (participant-level points or CIs) and to better emphasize within-feature deltas across conditions (e.g., dot-and-line per feature; small multiples). Add reference lines at $\mathrm{AUC}=0.5$ and optionally other benchmarks; use colorblind-safe encoding; ensure vector export and readable labels. Consider adding an explicit $\Delta\mathrm{AUC}$ panel (hip$\rightarrow$wrist, $100\rightarrow25$).
  • Related work and feature-set justification are thin and sometimes cite tangential domains rather than gait/step-counting literature (Sec. 1, Sec. 2.3, References). Novelty (systematic per-feature robustness across placement and sampling) is present but not crisply positioned.
    Recommendation: Add a focused related-work paragraph in Sec. 1 on prior findings about hip vs wrist and sampling-rate reduction for step counting, and clearly state the paper’s unique contribution. In Sec. 2.3, justify why these nine features are a representative low-cost set and briefly note omitted but relevant feature families (e.g., autocorrelation/periodicity measures). Tighten citations toward directly relevant accelerometry/gait sources.
  • Class imbalance and operational metrics: AUC is prevalence-insensitive, but deployment decisions often need operating-point metrics; imbalance may differ by condition (Sec. 2.4.1, Sec. 3.1).
    Recommendation: Report class prevalence per condition (Hip/$100$, Hip/$25$, Wrist/$100$, Wrist/$25$) in Sec. 3.1. Consider adding PR-AUC or sensitivity/specificity at fixed FPR (or fixed sensitivity) for the top features, to complement AUC and support “practical” conclusions (Sec. 3.6).
  • Ethics/data governance are not mentioned for human-participant data (Sec. 2.1).
    Recommendation: Add a brief ethics statement (IRB/ethics approval, consent, and data-sharing constraints) in Sec. 2 (new subsection if appropriate) following venue norms.
  • Placeholder/incorrect author–affiliation line and potentially irrelevant references reduce professionalism and may violate venue requirements (Title page; References).
    Recommendation: Replace the placeholder affiliation text with correct author/affiliation information (or proper anonymization for review). Audit the bibliography to remove/replace unrelated references and ensure each citation supports the corresponding claim.
Very Minor Issues:
  • Typographical/style inconsistencies (units “100Hz” vs “100~Hz”, “hipworn”, inconsistent feature naming/capitalization, quote styles for class labels, reference formatting) (Sec. 2.2.2, Sec. 3.x, Sec. 4, References).
    Recommendation: Proofread and standardize units, hyphenation, feature names/abbreviations (define once and reuse consistently), quote style for “Step/Non-Step,” and reference formatting per the target venue.
  • Section heading formatting inconsistency (e.g., stray leading “#” in a heading) (Sec. 3.3).
    Recommendation: Normalize heading formatting/numbering to match the manuscript’s overall style and the target venue template.
  • Dominant frequency definition may inadvertently select DC if ENMO is not mean-centered or if DC is not excluded (Sec. 2.3.2).
    Recommendation: State whether ENMO windows are demeaned and whether the DC bin is excluded; define the frequency search range used for “dominant frequency.”

Mathematical Consistency Audit

Mathematics Audit by Skepthical

This section audits symbolic/analytic mathematical consistency (algebra, derivations, dimensional/unit checks, definition consistency).

Maths relevance: light

The paper is primarily methodological/empirical with limited explicit mathematics. The main explicit formula is the ENMO transformation; other 'math' consists of standard descriptive-statistics feature definitions (mean/SD/variance/IQR/peaks) and high-level descriptions of FFT-derived features and AUC. There are no multi-step derivations to audit; the audit therefore focuses on dimensional consistency, unambiguous formulas, and definition consistency between sections.

Checked items

  1. ✖ ENMO definition (dimensional consistency and parentheses) (Introduction (p.2) and Sec. 2.2.1 (p.2–3))

    • Claim: ENMO is computed from triaxial acceleration as $\mathrm{ENMO} = \sqrt{a_x^2 + a_y^2 + a_z^2 - 1g}$, providing an orientation-invariant motion intensity centered around zero.
    • Checks: dimensional/units consistency, algebra/parentheses ambiguity, sanity case (stillness)
    • Verdict: FAIL; confidence: high; impact: critical
    • Assumptions/inputs: $a_x, a_y, a_z$ denote acceleration components with consistent units, $1g$ denotes gravitational acceleration
    • Notes: As written, $a_x^2+a_y^2+a_z^2$ has units of acceleration$^2$, but $1g$ has units of acceleration, so the subtraction under the square root is ill-typed. It also risks a negative radicand. To match the stated intent ('minus one g' after taking the Euclidean norm), the formula must be written unambiguously as $\mathrm{ENMO} = \sqrt{a_x^2+a_y^2+a_z^2} - 1g$ (optionally with truncation if intended).

  2. ✔ Window sample count at $100~\mathrm{Hz}$ (Sec. 2.2.2, p.3)

    • Claim: A $2$-second window corresponds to $200$ samples at $100~\mathrm{Hz}$.
    • Checks: arithmetic consistency
    • Verdict: PASS; confidence: high; impact: minor
    • Assumptions/inputs: Sampling frequency is exactly $100$ samples/second, Window duration is exactly $2$ seconds
    • Notes: $2~\mathrm{s} \times 100~\mathrm{Hz} = 200$ samples.

  3. ✔ Window sample count at $25~\mathrm{Hz}$ (Sec. 2.2.2, p.3)

    • Claim: A $2$-second window corresponds to $50$ samples at $25~\mathrm{Hz}$.
    • Checks: arithmetic consistency
    • Verdict: PASS; confidence: high; impact: minor
    • Assumptions/inputs: Sampling frequency is exactly $25$ samples/second, Window duration is exactly $2$ seconds
    • Notes: $2~\mathrm{s} \times 25~\mathrm{Hz} = 50$ samples.

  4. ✔ $50\%$ overlap implies $1$-second stride (Sec. 2.2.2, p.3)

    • Claim: Using $50\%$ overlap with a $2$-second window means each window advances by $1$ second.
    • Checks: definition consistency, arithmetic consistency
    • Verdict: PASS; confidence: high; impact: minor
    • Assumptions/inputs: Overlap is defined as fraction of window length shared between consecutive windows
    • Notes: Stride $= (1 - 0.5) \times 2~\mathrm{s} = 1~\mathrm{s}$.

  5. ✔ Variance definition equals SD squared (Sec. 2.3.1, p.3)

    • Claim: Signal variance is defined as the square of the standard deviation.
    • Checks: definition consistency
    • Verdict: PASS; confidence: high; impact: minor
    • Assumptions/inputs: Standard deviation is the (nonnegative) square root of variance
    • Notes: This is mathematically consistent as a definitional relationship (assuming population vs sample convention is consistently applied; the paper does not specify which).

  6. ✔ IQR definition (Sec. 2.3.1, p.3)

    • Claim: IQR is defined as the $75$th percentile minus the $25$th percentile of ENMO within the window.
    • Checks: definition consistency, sanity bounds
    • Verdict: PASS; confidence: high; impact: minor
    • Assumptions/inputs: Percentiles are computed on the set of window samples
    • Notes: $\mathrm{IQR} \geq 0$ by definition; consistent with use as dispersion.

  7. ⚠ Spectral energy definition vs 'power spectrum' wording (Sec. 2.3.2, p.3)

    • Claim: After FFT and obtaining a power spectrum, spectral energy ($0.5\text{--}3.0~\mathrm{Hz}$) is computed as the sum of squared magnitudes of FFT components in that band.
    • Checks: notation/definition consistency, dimensional consistency
    • Verdict: UNCERTAIN; confidence: medium; impact: moderate
    • Assumptions/inputs: Power spectrum may mean $|X[k]|^2$, FFT magnitude means $|X[k]|$
    • Notes: The text simultaneously references a 'power spectrum' and then 'squared magnitudes of FFT components'. If power spectrum is already $|X|^2$, then summing squared magnitudes would imply $|X|^4$. Likely the intended computation is $\sum |X|^2$ over the band, but this is not stated unambiguously.

  8. ⚠ Spectral entropy definition missing (Sec. 2.3.2, p.3)

    • Claim: Spectral entropy measures flatness/uniformity of the power spectrum; lower indicates concentrated energy.
    • Checks: missing definition/formula check
    • Verdict: UNCERTAIN; confidence: high; impact: moderate
    • Assumptions/inputs: Entropy requires a normalized distribution over frequency bins
    • Notes: No explicit entropy formula is provided (normalization, log base, frequency range, handling of zeros), preventing verification and comparability across conditions.

  9. ⚠ AUC interpretation vs feature directionality (Sec. 2.4.1, p.4; plus feature descriptions in Sec. 2.3.2, p.3)

    • Claim: AUC is the probability a random Step window is ranked higher (has a higher feature value) than a random Non-Step window.
    • Checks: definition consistency across sections, sign/direction sanity check
    • Verdict: UNCERTAIN; confidence: medium; impact: moderate
    • Assumptions/inputs: AUC computed from raw feature values with Step as positive class, Some features may be negatively associated with Step (e.g., entropy lower for steps)
    • Notes: The paper states lower spectral entropy corresponds to rhythmic walking, which would imply Step windows may have lower values than Non-Step. Under the stated AUC-as-'higher-is-more-positive' interpretation, such a feature would yield $\mathrm{AUC} < 0.5$ unless the direction is inverted. The paper does not specify direction handling.

  10. ⚠ Dominant frequency definition and DC handling (Sec. 2.3.2, p.3)

    • Claim: Dominant frequency is the frequency with the highest magnitude in the power spectrum.
    • Checks: definition completeness / edge cases
    • Verdict: UNCERTAIN; confidence: medium; impact: minor
    • Assumptions/inputs: FFT is applied to windowed ENMO data
    • Notes: If ENMO has a nonzero mean in a window, the $0~\mathrm{Hz}$ (DC) bin can dominate unless excluded or the signal is demeaned. The paper does not specify whether DC is excluded or whether demeaning/windowing is applied.

Limitations

  • Audit was performed on the provided parsed text of the $9$-page PDF; equation numbering, tables, and any mathematical details embedded solely in figures were not fully available for verification.
  • The paper contains few explicit formulas and no multi-step derivations; several signal-processing quantities (FFT scaling, power spectrum definition, entropy formula) are described only qualitatively, limiting the ability to verify mathematical equivalence across conditions without inventing missing steps.

Numerical Results Audit

Numerics Audit by Skepthical

This section audits numerical/empirical consistency: reported metrics, experimental design, baseline comparisons, statistical evidence, leakage risks, and reproducibility.

Out of $22$ numeric checks, $21$ passed and $1$ failed. The only failure is a cross-section inconsistency in total annotated step events between Methods and Results. All other checked arithmetic relationships (percentages, window counts, sampling/window math, and reported AUC drops/differences) match exactly or within stated tolerances.

Checked items

  1. ✔ C1 (p.2, Methods 2.1 Study Participants and Data Collection)

    • Claim: Cohort consisted of $19$ males ($48.7\%$) and $20$ females ($51.3\%$) out of $39$ participants.
    • Checks: percent_of_total_and_sum_to_100
    • Verdict: PASS
    • Notes: Percent checks and sums (counts and percents) all consistent within tolerance.

  2. ✔ C2 (p.2, Methods 2.1 Study Participants and Data Collection)

    • Claim: Age distribution: $25$ individuals ($64.1\%$) aged $18$–$39$; $14$ individuals ($35.9\%$) aged $40$–$65$ (total $39$).
    • Checks: percent_of_total_and_sum_to_100
    • Verdict: PASS
    • Notes: Percent checks and sums (counts and percents) all consistent within tolerance.

  3. ✔ C3 (p.4, Results 3.1 Data preparation and cohort summary)

    • Claim: $156$ data files correspond to $39$ participants across four experimental conditions.
    • Checks: product_equals_total
    • Verdict: PASS
    • Notes: $39 \times 4 = 156$ exactly.

  4. ✔ C4 (p.4, Results 3.1 Data preparation and cohort summary)

    • Claim: Total windows $545,350$ equals Step windows $157,710$ plus Non-Step windows $387,640$.
    • Checks: parts_sum_to_total
    • Verdict: PASS
    • Notes: $157,710 + 387,640 = 545,350$ exactly.

  5. ✖ C5 (p.2, Methods 2.1 vs p.4, Results 3.1)

    • Claim: Total annotated step events reported as $48,721$ (Methods) vs $62,904$ (Results).
    • Checks: cross_section_numeric_consistency
    • Verdict: FAIL
    • Notes: Mismatch detected: Results minus Methods $= 14,183$; equality would be expected if same scope/definition.

  6. ✔ C6 (p.3, Methods 2.2.2 Sliding Window Segmentation and Labeling)

    • Claim: $2$-second window corresponds to $200$ samples at $100~\mathrm{Hz}$ and $50$ samples at $25~\mathrm{Hz}$.
    • Checks: unit_rate_times_duration
    • Verdict: PASS
    • Notes: $100 \times 2 = 200$ and $25 \times 2 = 50$ exactly.

  7. ✔ C7 (p.3, Methods 2.2.2)

    • Claim: $50\%$ overlap implies $1$-second stride for a $2$-second window.
    • Checks: overlap_stride_consistency
    • Verdict: PASS
    • Notes: Stride computed as $2 \times (1-0.5) = 1$ exactly.

  8. ✔ C8 (p.6, Results 3.4 (hip sampling frequency robustness))

    • Claim: Hip: std_dev and variance AUC dropped by $0.0036$ (from $0.9839$ at Hip/$100~\mathrm{Hz}$ to $0.9803$ at Hip/$25~\mathrm{Hz}$).
    • Checks: difference_equals_reported_drop
    • Verdict: PASS
    • Notes: $0.9839-0.9803$ matches reported drop within tolerance.

  9. ✔ C9 (p.5, Results 3.3 (location robustness at $100~\mathrm{Hz}$))

    • Claim: At $100~\mathrm{Hz}$: std_dev and variance AUC decreased by $0.0704$ (from $0.9839$ at Hip/$100~\mathrm{Hz}$ to $0.9135$ at Wrist/$100~\mathrm{Hz}$).
    • Checks: difference_equals_reported_drop
    • Verdict: PASS
    • Notes: $0.9839-0.9135$ matches reported drop within tolerance.

  10. ✔ C10 (p.5, Results 3.3 (location robustness at $100~\mathrm{Hz}$))

    • Claim: At $100~\mathrm{Hz}$: spectral_energy drop of $0.0509$ (from $0.9839$ to $0.9330$).
    • Checks: difference_equals_reported_drop
    • Verdict: PASS
    • Notes: $0.9839-0.9330$ matches reported drop within tolerance.

  11. ✔ C11 (p.5, Results 3.3 (IQR location robustness at $100~\mathrm{Hz}$))

    • Claim: IQR AUC decreased by $0.0395$ at $100~\mathrm{Hz}$ ($0.9817$ at Hip/$100~\mathrm{Hz}$ to $0.9422$ at Wrist/$100~\mathrm{Hz}$).
    • Checks: difference_equals_reported_drop
    • Verdict: PASS
    • Notes: $0.9817-0.9422$ matches reported drop within tolerance.

  12. ✔ C12 (p.5, Results 3.3 (IQR location robustness at $25~\mathrm{Hz}$))

    • Claim: IQR AUC decreased by $0.0357$ at $25~\mathrm{Hz}$ ($0.9771$ at Hip/$25~\mathrm{Hz}$ to $0.9414$ at Wrist/$25~\mathrm{Hz}$).
    • Checks: difference_equals_reported_drop
    • Verdict: PASS
    • Notes: $0.9771-0.9414$ matches reported drop within tolerance.

  13. ✔ C13 (p.5, Results 3.3 (peak_count location degradation at $100~\mathrm{Hz}$))

    • Claim: peak_count AUC plummeted by $0.2809$ (from $0.9598$ at Hip/$100~\mathrm{Hz}$ to $0.6789$ at Wrist/$100~\mathrm{Hz}$).
    • Checks: difference_equals_reported_drop
    • Verdict: PASS
    • Notes: $0.9598-0.6789$ matches reported drop within tolerance.

  14. ✔ C14 (p.6, Results 3.4 (wrist sampling frequency robustness: peak_count increase))

    • Claim: peak_count at wrist increased in AUC from $0.6789$ ($100~\mathrm{Hz}$) to $0.7658$ ($25~\mathrm{Hz}$).
    • Checks: direction_and_difference
    • Verdict: PASS
    • Notes: Direction check passed ($0.7658 > 0.6789$); computed increase $= 0.0869$.

  15. ✔ C15 (p.6, Results 3.4 (hip sampling frequency robustness: dominant_freq increase))

    • Claim: dominant_freq at hip increased from $0.6500$ ($100~\mathrm{Hz}$) to $0.7317$ ($25~\mathrm{Hz}$).
    • Checks: direction_and_difference
    • Verdict: PASS
    • Notes: Direction check passed ($0.7317 > 0.6500$); computed increase $= 0.0817$.

  16. ✔ C16 (p.5, Results 3.2 (dominant_freq examples))

    • Claim: dominant_freq yielded AUCs $0.5265$ at Wrist/$100~\mathrm{Hz}$ and $0.7317$ at Hip/$25~\mathrm{Hz}$ (examples).
    • Checks: range_and_consistency_with_random_chance_reference
    • Verdict: PASS
    • Notes: Both AUCs within $[0,1]$; wrist/$100~\mathrm{Hz}$ above $0.5$ by $0.0265$.

  17. ✔ C17 (p.5, Results 3.2 (spectral_entropy range statement))

    • Claim: spectral_entropy AUCs ranged from $0.5136$ to $0.8474$.
    • Checks: range_ordering_and_bounds
    • Verdict: PASS
    • Notes: Ordering and bounds check passed: $0 \leq 0.5136 \leq 0.8474 \leq 1$.

  18. ✔ C18 (p.6, Results 3.4 (spectral_entropy frequency reduction, hip))

    • Claim: spectral_entropy AUC at hip dropped from $0.7210$ ($100~\mathrm{Hz}$) to $0.5136$ ($25~\mathrm{Hz}$).
    • Checks: difference_and_direction
    • Verdict: PASS
    • Notes: Direction check passed ($0.7210 > 0.5136$); computed drop $= 0.2074$.

  19. ✔ C19 (p.6, Results 3.4 (spectral_entropy frequency reduction, wrist))

    • Claim: spectral_entropy AUC at wrist dropped from $0.8474$ ($100~\mathrm{Hz}$) to $0.7164$ ($25~\mathrm{Hz}$).
    • Checks: difference_and_direction
    • Verdict: PASS
    • Notes: Direction check passed ($0.8474 > 0.7164$); computed drop $= 0.1310$.

  20. ✔ C20 (p.6, Results 3.4 (mean feature hip drop))

    • Claim: Mean AUC at hip dropped from $0.8488$ ($100~\mathrm{Hz}$) to $0.7293$ ($25~\mathrm{Hz}$).
    • Checks: difference_and_direction
    • Verdict: PASS
    • Notes: Direction check passed ($0.8488 > 0.7293$); computed drop $= 0.1195$.

  21. ✔ C21 (p.6, Results 3.4 (wrist sampling frequency robustness: iqr difference))

    • Claim: IQR at wrist had difference of $0.0008$ between $100~\mathrm{Hz}$ and $25~\mathrm{Hz}$ (AUCs $0.9422$ and $0.9414$).
    • Checks: difference_equals_reported
    • Verdict: PASS
    • Notes: $|0.9422-0.9414|$ matches reported $0.0008$ within tolerance.

  22. ✔ C22 (p.5-7, Narrative claims about thresholds)

    • Claim: Claim: variability/energy features achieved high AUCs (all $>0.91$) across all conditions; and later: IQR (AUC $> 0.94$) and spectral_energy (AUC $> 0.92$) for wrist data.
    • Checks: threshold_claim_vs_listed_values
    • Verdict: PASS
    • Notes: For explicitly listed wrist values: IQR $0.9422$ and $0.9414$ exceed $0.94$; spectral_energy $0.9330$ exceeds $0.92$.

Limitations

  • Only the provided PDF text was used; Table 1 and Table 2 numeric contents were not available in the parsed text, preventing verification of claims that require the full table.
  • No values were extracted from plots/figures (no pixel/bar reading), so any claim relying solely on Figure 1 without explicit numbers in the text is not directly checkable.
  • Many aggregate quantities (e.g., total windows from recording durations) cannot be recomputed without per-participant durations and preprocessing rules not explicitly specified.
  • The detected mismatch in annotated step-event totals ($48,721$ vs $62,904$) cannot be resolved to a single correct value without additional clarification about definitions/scope.