A reproducible processed-feature LOSO benchmark for sEMG protocol-state discriminability in a cable-driven walking-assistance suit

A reproducible processed-feature LOSO benchmark for sEMG protocol-state discriminability in a cable-driven walking-assistance suit | Research Square window.SnipcartSettings = { analytics: { enabled: false } }; (function() { var accessVector = localStorage.getItem('access_vector') || ''; window.dataLayer = window.dataLayer || []; if (accessVector) { window.dataLayer.push({ user: { profile: { profileInfo: { snid: accessVector } } } }); } })(); (function(w,d,s,l,i){w[l]=w[l]||[];w[l].push({'gtm.start':new Date().getTime(),event:'gtm.js'});var f=d.getElementsByTagName(s)[0],j=d.createElement(s),dl=l!='dataLayer'?'&l='+l:'';j.async=true;j.src='https://www.googletagmanager.com/gtm.js?id='+i+dl;f.parentNode.insertBefore(j,f);})(window,document,'script','dataLayer','GTM-K279D39R'); Browse Preprints In Review Journals COVID-19 Preprints AJE Video Bytes Research Tools Research Promotion AJE Professional Editing AJE Rubriq About Preprint Platform In Review Editorial Policies Our Team Advisory Board Help Center Sign In Submit a Preprint Cite Share Download PDF Research Article A reproducible processed-feature LOSO benchmark for sEMG protocol-state discriminability in a cable-driven walking-assistance suit Song-Bi Lee¹, Yongjun Kim¹, Gisu Heo¹, Changmok Oh¹, Suyeong Eom¹, and 1 more This is a preprint; it has not been peer reviewed by a journal. https://doi.org/ 10.21203/rs.3.rs-9590227/v1 This work is licensed under a CC BY 4.0 License Status: Under Review Version 1 posted 8 You are reading this latest preprint version Abstract Background We present a null-referenced processed-feature leave-one-subject-out (LOSO) transfer-boundary benchmark for surface electromyography (sEMG) protocol-state discriminability in healthy adult men using a cable-driven walking-assistance suit. Protocol states are defined as not-worn (NW), worn-inactive suit (UE), and worn-active suit (PE). PE denotes active assistance in the same worn hardware configuration as UE; actuator commands, cable-tension traces, and torque profiles were excluded from predictors. Methods Data from 10 healthy adult men were analyzed. The processed dataset comprised 29,993 EMG-derived gait cycles (239,944 cycle×phase samples; 10-channel sEMG at 2,148 Hz). We defined a fixed subject-independent LOSO benchmark with a balanced manifest capped at ≤ 50 cycles per recording file, cycle-level macro-F1 as the primary endpoint, separated calibration regimes, and offline computational-cost readouts. Leakage control denotes subject-level split separation and exclusion of held-out-subject labels and normalization statistics, not removal of residual order, fatigue, electrode-state, or protocol-speed confounding. The reproducible unit is the processed-feature recognition benchmark; raw EMG preprocessing and EMG-derived event extraction are documented but not fully rerunnable from the public package because raw time series are available only under DUA. Results Cycle-level macro-F1 was 0.41–0.43 within condition and 0.18–0.41 across protocol-condition transfer cells. Direct cycle-level Random Forest and multinomial logistic-regression comparators preserved the same diagonal-versus-off-diagonal ordering. Offline computational-cost readouts separated a high-cost reference configuration from a lower-latency comparator (inference-only mean 1.36 ms; lower-latency end-to-end mean 115 ms; high-cost end-to-end mean 2,697 ms). The lower-latency comparator was not presented as a deployment-ready control claim. Uncertainty was estimated at the held-out-subject level; cycle and phase counts should not be interpreted as independent participant counts. Conclusions The contribution is a fixed benchmark procedure—split manifest, transfer grid, null references, scoring endpoint, and reproducibility package—rather than a new classifier or deployment controller. The modest macro-F1 values are part of the benchmark finding: under fixed leakage-controlled LOSO scoring, processed-feature EMG-only recognition showed limited subject-independent transfer across protocol-condition bundles. Findings should be interpreted as protocol-condition transfer across bundled terrain/platform/speed conditions in this healthy adult male proof-of-concept cohort. Trial registration Not applicable. This study is a recognition-stage methodological benchmark and does not evaluate prospectively assigned health-care interventions against clinical health outcomes. Trial registration Not applicable. Cable-driven suit surface electromyography protocol-state discriminability subject-independent benchmark leave-one-subject-out calibration computational latency reproducibility Figures Figure 1 Figure 2 Figure 3 Figure 4 Background Wearable gait-assistance systems such as cable-driven suits may require real-time discrimination of user assistance states for adaptive closed-loop control [ 1 – 7 ]. Surface electromyography (sEMG) can non-invasively measure muscle activity and has been used for gait-mode identification, prosthetic control, and classification [ 8 – 14 ]. A wide variety of time-domain, frequency-domain, and wavelet features have been proposed for EMG classification [ 9 , 15 – 26 ], and nonlinear dynamics measures can capture signal complexity associated with gait variability [ 27 – 37 ]. However, the multi-terrain assistance-state sEMG classification field has not converged on a shared definition of evaluation units, allowed test-time information, and risk/latency metrics, so a single number — classifier accuracy = X% — is rarely sufficient to compare two systems under matched conditions. Despite this rich methodological landscape, subject-independent sEMG-based protocol-state discriminability still lacks an agreed evaluation grid that specifies which subjects, which terrain/platform/speed bundle conditions, and which test-time information are allowed for comparison. This proof-of-concept study defines and evaluates a leakage-controlled recognition-stage benchmark for sEMG-based discriminability of externally imposed NW/UE/PE protocol states in a single cable-driven suit platform. The primary objective is to quantify subject-independent cycle-level macro-F1 under within-condition LOSO and across protocol-defined terrain/platform/speed bundle transfer cells. Secondary objectives are to document regime-specific calibration behavior and computational latency constraints, without treating these analyses as deployment-readiness evidence. To support reproducible comparison, we release a fixed split manifest, scoring script, fixed seeds, environment lock, and processed-feature package. Thus, the methodological contribution is a bounded benchmark protocol rather than a new classifier architecture: it specifies the evaluation grid, allowed test-time information, null references, calibration regimes, and cycle-level scoring endpoint needed for reproducible comparison of future recognition methods. Operational rationale of the benchmark label. The purpose is not to infer a suit controller state that the system already knows, but to quantify whether wearable-hardware presence and active-assistance protocol leave measurable subject-independent signatures in the sEMG feature space under leakage-controlled evaluation. Accordingly, NW/UE/PE are used as protocol-state discriminability labels for benchmark comparison, not as a substitute for controller-state telemetry. Relevance to JNER readership. Quantifying when subject-independent transfer fails is practically important for neuroengineering and rehabilitation-assistance pipelines because it defines the reliability boundary of EMG-only recognition modules before they are coupled to adaptive control or clinical translation workflows. Methods Phase-training and cycle-level endpoint boundary. The model was trained on phase-bin samples, but all primary performance claims use one prediction per gait cycle. Because phase bins within a cycle are correlated, uncertainty was estimated only at the held-out-subject level. Direct cycle-level RF and logistic-regression comparators were included to check that the primary transfer ordering was not an artifact of phase-level training. Participants Ten healthy adult men (age 25.3 ± 3.2 years; height 168.4 ± 8.7 cm; weight 64.2 ± 11.3 kg) participated. The study was approved by the Institutional Review Board (IRB) of the Electronics and Telecommunications Research Institute (ETRI) under approval number N01-202409-01-020, and written informed consent was obtained from every participant. No a priori sample-size calculation was performed because this was a proof-of-concept benchmark based on the available cable-driven suit dataset; all eligible participants with complete NW/UE/PE recordings across the three protocol-defined walking conditions were included. Cable-driven suit and suit states In this benchmark, NW, UE, and PE are externally imposed protocol-state labels. NW denotes walking without the suit, UE denotes the suit worn with assistance disabled, and PE denotes the same worn configuration with active assistance enabled under a fixed gait-synchronous study controller. UE and PE therefore differ at the protocol level by actuator active/inactive status, not by suit placement or wearable configuration. The classifier is not intended to infer a controller state unknown to the device, estimate torque magnitude, reconstruct actuator commands, or decode user intent. Instead, the labels are used to test whether wearable-hardware presence and active-assistance protocol leave subject-independent signatures in the processed sEMG feature space. The released benchmark package supports reproduction of the recognition-stage endpoint, not reproduction of the actuator-control waveform. Controller-command traces, continuous force/torque time series, internal controller states, update-rate details, and trigger thresholds are excluded from the released predictors and scoring inputs. Minimum non-identifying protocol metadata are provided to define the PE-versus-UE contrast at the benchmark level. Minimum PE descriptor and release boundary. UE and PE used the same wearable hardware configuration: a cable-driven soft suit with a back-mounted actuator unit and strap-fixed thigh/pelvic anchor interfaces. In PE, four motor-driven cable channels provided gait-synchronous bilateral hip flexion/extension-related assistance through the anchor path; in UE, assistance was disabled under the same worn configuration and fixed study-controller settings. No subject-specific controller retuning or hip-joint-center reconstruction was used. The released predictors include only processed sEMG features and protocol metadata; actuator commands, cable-tension/force–torque traces, and internal controller states were excluded. Therefore, PE is interpreted as a categorical active-assistance protocol state, and detailed release-boundary descriptors are provided in Supplementary Table S0 and Additional file 1. Minimum non-identifying PE protocol metadata and the benchmark release boundary are summarized in Supplementary Table S0. Walking protocol and protocol-condition definition Data were collected under three suit states (NW, UE, PE) and three protocol-defined walking conditions. Level walking was performed on a treadmill at 0.83 m/s. Stair walking was performed on an Egojin Highclimb stepmill (step height 21 cm) at 0.54 m/s. Slope walking was performed on a 25 m walkway at an 18° gradient at self-selected speed (50 m round trip; trial-mean speed 0.227 ± 0.016 m/s). Mean recording durations were 299 ± 23 s (level), 221 ± 17 s (slope), and 180 ± 6 s (stairs). Each 18° slope round trip included two opposite-direction segments at the same gradient, which were treated as uphill and downhill in analysis. Uphill and downhill segments were retained within the single slope protocol-condition bundle for the primary benchmark rather than analyzed as separate primary conditions. Protocol metadata included subject, protocol condition, suit state, and trial/run identifiers; randomization/counterbalancing order, scheduled rest intervals, and electrode reattachment logs were not explicitly recorded in the released benchmark metadata. Protocol-condition metadata are summarized in Additional file 1 Table A2. Throughout this paper, “protocol condition” denotes the (terrain + platform + protocol-mean speed) bundle defined above as a single evaluation unit. Terrain and protocol speed are therefore jointly varied; full “speed-only” decoupling is out of scope. Additional file 1 provides the supplementary index mapping all Supplementary Figures and Tables to file locations. Additional files 2–8 provide, respectively, domain-level permutation importance, transfer-cell source tables, condition-label mismatch sensitivity, calibration source tables, stride-time-matched speed-confound control, subject-level/biomechanical/comparator tables, and event/F18-F28 sensitivity plus endpoint-hierarchy and empirical-null robustness analyses. Partial decoupling via stride-time matching is reported in Additional file 6. Box 1. Minimum in-body PE active-assistance descriptor. Field In-body PE descriptor Wearable platform Cable-driven soft suit with a back-mounted actuator unit and strap-fixed thigh/pelvic anchor interfaces. UE-vs-PE protocol contrast Same worn hardware in both states; PE = actuator assistance enabled, UE = assistance disabled. Assistance target and direction Bilateral hip flexion/extension-related assistance under fixed gait-synchronous timing. Anchor path convention Motor-driven cable lines were routed through strap-fixed body-surface anchor regions (force-transfer and routing/guide functions). Public benchmark predictors Processed sEMG features and protocol metadata only. Interpretation boundary PE is a categorical active-assistance protocol state; actuator-level and biomechanical detail remains outside the public benchmark scope (see Supplementary Table S0 / Additional file 1). Box 1 summarizes only the minimum in-body PE protocol descriptor used for main-text interpretation. The full release-boundary and non-claim detail is provided in Supplementary Table S0 and Additional file 1. EMG acquisition Ten-channel sEMG was collected from gastrocnemius medialis (GM), tibialis anterior (TA), rectus femoris (RF), tensor fasciae latae (TFL), and biceps femoris (BF) bilaterally with the Delsys Trigno surface EMG system at 2,148 Hz, following SENIAM electrode-placement guidelines [ 38 ]. Preprocessing Each recording was preprocessed per channel with the following steps applied in order: non-zero-sample-based outlier clipping at the 0.1–99.9 percentiles (computed only when ≥ 100 non-zero samples were available; zeros were excluded from the percentile statistics but retained in the signal), a 4th-order Butterworth band-pass filter (20–450 Hz), a 60 Hz notch filter (Q = 30), full-wave rectification, and Gaussian envelope smoothing (σ = 100 ms). Gait segmentation Throughout this manuscript we use EMG-derived HS/TO proxies rather than externally validated gait events. An EMG-derived heel-strike (HS) proxy was detected as a TA-envelope peak; an EMG-derived toe-off (TO) proxy was detected as a GM-envelope peak (scipy.signal.find_peaks; minimum distance 0.5 s). For peak-detection robustness, envelopes were normalized by median/MAD and peaks were searched stepwise with relaxed (height, prominence) = (0.5, 0.3) → (0.3, 0.2) → (0.1, 0.1) → (0.05, 0.05) → (None, 0.02) (terminating at ≥ 5 peaks); the final pass used minimum distance only. A side with fewer than 2 HS peaks was excluded. HS/TO and HS→next-HS cycles were constructed independently for left and right; per-side cycles were then pooled. For each cycle, TO was selected as the GM peak nearest 62% of the cycle length after HS; if absent, TO was estimated at 62%. Cycles were divided into eight fixed-percentage cycle bins based on Perry timing references [ 39 ]: IC (0–2%), LR (2–12%), MSt (12–31%), TSt (31–50%), PSw (50–62%), ISw (62–75%), MSw (75–87%), and TSw (87–100%). These bins are fixed-percentage cycle bins within each HS→next-HS cycle and are not externally validated biomechanical gait phases. Phase boundaries used Perry’s fixed percentages and were not shifted by the detected TO. The detected TO proxy was retained only as a reference event for cycle characterization and robustness checks; the fixed-percentage bins used for classification were assigned by percentages within each HS→next-HS cycle and were mapped to sample indices within that interval (no time resampling). Cycles shorter than 50 samples were excluded. Because gait-cycle boundaries and fixed-percentage phase bins were derived from EMG envelopes rather than external kinematic/force-plate events, this benchmark evaluates recognition under an EMG-derived segmentation convention. It should not be interpreted as a validation of biomechanical gait-phase detection. Event-proxy QC summary is provided in Additional file 8: sides failing minimum-peak criteria were excluded before classification, retained cycle durations were constrained by the cycle-construction filters described above, and plausible event/phase perturbation sensitivity checks did not change the primary transfer-cell ordering. Across the full collected dataset after cycle-construction filters, 29,993 cycles were retained for downstream benchmark analyses (Table 1 a). Table 1 Dataset distribution and primary cycle-level LOSO benchmark. a) Gait-cycle distribution by protocol condition and suit state (full collected data). Protocol condition NW UE PE Total Level 5583 5565 5307 16455 Slope (18°) 2034 2124 2254 6412 Stairs 2290 2409 2427 7126 Total 9907 10098 9988 29993 NW, not worn; UE, worn-inactive suit; PE, worn-active suit. The three protocol-defined walking conditions were performed on different platforms (only the level condition used a treadmill): level treadmill, 0.83 m/s; 18° slope on a 25 m walkway (50 m round trip, self-selected; trial mean 0.227 m/s); stairs on an Egojin Highclimb stepmill (step height 21 cm), 0.54 m/s. This sub-table reports the full collected distribution (29,993 cycles total) and is independent of the balanced LOSO evaluation set (≤ 50 cycles per recording file). b) Cycle-level (primary endpoint): 3 × 3 protocol-condition transfer matrix, where each condition bundles terrain, platform, and protocol speed (LOSO macro-F1; mean ± 95% subject-bootstrap CI; n = 10). Train ↓ \ Test → Level (treadmill, 0.83 m/s) Slope (walkway 18°, 0.227 m/s) Stairs (stepmill, 0.54 m/s) Level 0.43 [0.34, 0.51] 0.21 [0.15, 0.28] 0.18 [0.12, 0.25] Slope 0.24 [0.17, 0.32] 0.41 [0.33, 0.49] 0.41 [0.33, 0.49] Stairs 0.23 [0.17, 0.30] 0.37 [0.30, 0.45] 0.42 [0.34, 0.50] Cycle-level macro-F1 = mean of available retained fixed-percentage bin prediction probabilities of one HS→next-HS cycle (after NL10 filtering), then argmax. Diagonal = within-condition; off-diagonal = protocol-defined terrain/platform/speed bundle transfer. This table is the primary endpoint because the operational decision unit is one decision per gait cycle. CIs are 2,000 subject bootstraps; per-subject raw values and cell-wise confidence intervals are provided in Additional file 3. Analytical references for interpretation are class-prior-matched random ≈ 0.333 and global-majority ≈ 0.168; empirical dummy/permutation references are reported in Supplementary Table S14 and used for null interpretation. Table 2 Evaluation regimes (main-text summary). Regime Allowed test-time information Endpoint Role 1. Within-condition LOSO none cycle-level macro-F1 Primary subject-independent reference within one protocol condition 2. 3×3 protocol-condition transfer (LOSO) none cycle-level macro-F1 (with cycle×phase sensitivity readout) Primary transfer-boundary reference across bundled protocol conditions 3. Known-condition reference normalization protocol-condition label only (no held-out suit-state labels) cycle×phase delta macro-F1 / ratio change versus Regime 2 Secondary known-condition calibration sensitivity 4. Offline pooled subject-window normalization first 4 unlabeled cycles per held-out subject and condition cycle×phase macro-F1 on remaining cycles Secondary unlabeled adaptation sensitivity (offline reference) 5. Small-label post-hoc calibration (LOTO under LOSO) small labeled cycle budget (phase-label-equivalent reporting) cycle×phase macro-F1 Secondary label-budget recovery characterization (different regime) 6. Latency benchmark n/a inference-only and end-to-end latency, reported separately Computational-cost readout (not deployment-readiness claim) 7. Asymmetric critical-error readout n/a P(pred = PE|true = NW) and P(pred = NW|true = PE) Risk visibility beyond macro-F1 within the same regime Full regime definitions and supplementary index mapping are provided in Additional file 1. Feature extraction F18 (18-D) consisted of time-domain (6), frequency-domain (5), and wavelet-domain (7) features. Event detection was performed on the envelope; F18 features were computed on the band-passed signal (bp) and its envelope (env): RMS / MAV / variance from env; waveform length (WL), zero-crossing rate (ZCR), slope sign change (SSC) from bp (sign-change-based, no ε threshold). Frequency features used Welch’s PSD on bp (scipy.signal.welch; nperseg = min(L, 1024); other parameters at SciPy defaults), restricted to 20–450 Hz; relative band power was defined as: integral of P(f) over the target band divided by integral of P(f) over 20–450 Hz. To avoid target-label dependence, wavelet features used a condition-agnostic fixed 3-wavelet ensemble (level = 5): {db5, sym6, coif3} with fixed weights {0.5, 0.3, 0.2}. NL10 (10-D) included sample entropy (m = 2, r = 0.2; z-scored signal), approximate entropy (m = 2, r = 0.2), permutation entropy (order = 3, delay = 1, normalized), Higuchi FD (kmax = min(10, N/4)), Katz FD, Hurst (max_lag = min(50, N/4)), recurrence-quantification metrics (m = 2, τ = 1, threshold = 0.1× max distance; DET lmin = 2; LAM vmin = 2), and the Rosenstein largest Lyapunov exponent (m = 2, τ = 1). NL10 was computed only on segments ≥ 50 samples (≈ 23 ms at 2,148 Hz); shorter segments were NaN, and any cycle×phase sample with NaN was excluded from F28 analyses (Supplementary Table S3 ). Feature-block definitions are summarized in Additional file 1 Table A3. Wavelet 7-D terms were computed from the same level-5 decomposition for each basis w in {db5, sym6, coif3}: WE_w = sum_k c_w,k^2 (detail-coefficient energy), WEnt_w = -sum_k p_w,k log p_w,k with p_w,k = c_w,k^2 / sum_j c_w,j^2, and cross-wavelet ratio = (WE_db5 + WE_sym6) / (WE_coif3 + 1e-12). These descriptors are reported as computational features for classification and not as direct mechanistic biomarkers. Predefined literature-based per-muscle scaling convention Per-muscle scaling weights for muscle m were defined as follows: w_raw,m = w_PCSA,m x w_FT,m x w_pen,m; w_PCSA,m = 1 / PCSA_m; w_FT,m = 1 + 0.6 x (FT_m − 0.5); w_pen,m = 1 - theta_m / 90 deg; w_m = w_raw,m / mean(w_raw across muscles); Phi_B = diag(w). where PCSA is physiological cross-sectional area, FT is fast-fibre fraction, and theta is pennation angle. The normalization sets the across-muscle mean of w to 1.0. Phi_B is a diagonal per-muscle scalar applied to all extracted features and is not personalized per subject. Muscle parameters (PCSA, FT, theta) were taken from the anatomical literature [ 40 , 41 ] as fixed per-muscle source-value triplets (left = right), with the resulting Phi_B weights reported in Supplementary Table S7 . Phi_B was retained only as a frozen, label-independent benchmark preprocessing convention, not because it improved performance or because it represents a validated subject-specific biomechanical model. Because Phi_B did not show a uniform performance gain, all primary transfer-boundary conclusions were checked against Phi_B-off comparators, and no claim depends on Phi_B improving recognition performance. No-leakage design summary • One analysis sample is one cycle×phase 28-D vector (10-channel mean after per-muscle scaling). • Primary endpoint is cycle-level macro-F1 (mean of available retained phase probabilities per HS→next-HS cycle). • LOSO split holds out one subject entirely; no held-out-subject labels are used in model fitting. • In this manuscript, leakage-controlled means that held-out-subject suit-state labels, held-out-subject class information, and held-out-subject feature statistics were not used for model fitting, normalization, calibration, feature scaling, or prediction. Ground-truth labels were used only after prediction for pre-specified evaluation-set definition and metric computation. • Signal-level preprocessing and EMG-event proxy construction are offline within-recording feature-generation steps, not causal online deployment procedures. • Balanced LOSO evaluation uses ≤ 50 cycles per recording file (recording file = one subject-specific trial/run under one protocol condition and one suit state) to avoid dominance by long files. • Known-condition reference normalization uses training-subject-only (mu, sigma) by protocol-defined condition and requires the correct test protocol-condition label. • The pooled subject-window analysis was retained as an offline reference normalization sensitivity analysis because the calibration window depended on recording-file structure and was excluded from evaluation. • Class-stratified post-hoc calibration characterization samples labels at whole-cycle level and excludes the sampled cycles in full. Classification The benchmark hierarchy was defined before reporting: cycle-level macro-F1 under LOSO was the primary endpoint; all cycle×phase, calibration, latency, coordination, and ablation analyses were secondary or supporting readouts. A Random Forest classifier (RF; scikit-learn 1.7.1 [ 42 ]; n_estimators = 200, random_state = 42) [ 43 ] was used under LOSO with a three-class target (NW, UE, PE). Each cycle×phase sample was a single 28-D vector — the 10-channel arithmetic mean of the per-channel F28 features (F28 = F18 + NL10) after the frozen per-muscle preprocessing convention (Φ_B), treated as a predefined label-independent convention rather than a learned performance-enhancement step. The cycle-level prediction is the mean of the available retained phase probabilities within one HS→next-HS cycle and is the primary endpoint; cycle×phase numbers are reported only as a finer-grained sensitivity readout. Training unit vs. endpoint unit. The primary model is trained on phase-level samples (8 phase bins per cycle), but the operational decision unit and primary endpoint are at the cycle level (one decision per HS→next-HS cycle). We retain phase-level training because the predefined benchmark implementation aggregates phase-wise class probabilities to one cycle-level decision while preserving phase-resolved sensitivity readouts (Fig. 2 b). Because phase samples within one cycle are correlated, primary uncertainty is computed at held-out subject level (subject bootstrap). A cycle-weighted training sensitivity (sample weights of 1 / retained-phase-count per cycle) is reported in Additional file 8 (Supplementary Table S12). As an additional robustness analysis, direct cycle-level comparators were evaluated on the same split manifest by aggregating retained phase features within each cycle using mean pooling (cycle-level 28-D vectors), followed by cycle-level RF and multinomial logistic regression under LOSO (Supplementary Table S13, Additional file 8). This design does not claim phase-level training superiority; it defines the benchmark hierarchy and its cycle-level comparators explicitly. Accordingly, phase-level training is treated as an implementation choice of the benchmark baseline, whereas all primary interpretation is tied to the cycle-level endpoint and to robustness checks that remove or down-weight within-cycle phase multiplicity. Balanced LOSO sampling algorithm (seed = 42). 1. Build a file list where one recording file is one (subject, protocol condition, suit state, trial/run file) unit. 2. For each file, select up to 50 HS→HS cycles once using seed 42 and write selected cycle IDs to splits/loso_manifest_v1.csv. 3. Reuse the same manifest across all LOSO folds (no per-fold re-sampling). 4. Train/evaluate only on manifest-selected cycles (capped design for both training and evaluation in this benchmark definition). 5. After NL10 filtering, retain a cycle for cycle-level scoring if at least one phase probability remains; compute cycle-level probabilities as the arithmetic mean of retained phase probabilities. For F28 analyses, cycle-level aggregation was computed from retained phase-level F28 entries after NL10 filtering (Supplementary Table S3 ); cycle-level denominators therefore correspond to cycles with available phase probabilities under this filter. Secondary calibration analyses Known-condition reference normalization. Condition-specific z-score normalization was computed from training subjects only (train_only_by_condition). For each feature dimension x, x’ = (x − µ_{train, condition}) / σ_{train, condition}; µ and σ were estimated from training subjects only and selected according to the protocol-defined condition label of each sample. No held-out-subject statistics or labels were used. For a train-condition → test-protocol-condition transfer cell, classifier fitting used only train-condition samples from training subjects, whereas test-condition normalization statistics were estimated from training-subject recordings in that same test condition. This regime is therefore a known-condition reference setting, not a condition-agnostic unsupervised adaptation setting. Offline pooled subject-window normalization analysis. In the held-out subject, the first 4 cycles per recording file per test condition were separated as an unlabelled calibration window and excluded from evaluation. Calibration-window cycles were pooled across recording files within each held-out subject × test condition cell to estimate one normalization vector. No suit-state-specific or file-specific normalization statistics were estimated or applied. Because the window is defined by recording-file structure, this analysis is retained as an offline reference normalization sensitivity analysis rather than an autonomous deployment policy. Class-stratified post-hoc calibration characterization. In leave-one-protocol-condition-out under LOSO, two protocol-defined conditions were used for training and the held-out condition for testing. Known-condition reference normalization was applied first, and a multinomial logistic regression was then fitted post-hoc to log-RF-probability outputs in the held-out subject × test condition. Labels were sampled and excluded at the whole-cycle level. Budgets are reported as phase-label-equivalent units (10/30/90), and the procedure was repeated 30 times with random cycle sampling (seed = 42). This LOTO-under-LOSO regime is a secondary characterization only and is excluded from primary-endpoint effectiveness claims. Uncertainty estimation and supporting statistical analyses 95% confidence intervals were obtained by 2,000 subject bootstraps (resampling subjects with replacement; 2.5–97.5 percentile range). Macro-F1 was computed per held-out subject and aggregated as the unweighted (subject) mean; CIs were computed on these subject-level scores. For paired regime-comparison deltas, Δ macro-F1 was computed first within held-out subject, and 95% CIs were obtained by bootstrapping the paired subject-level differences with the same 2,000 subject-resampling procedure. Domain-level group permutation importance was computed under in-condition LOSO (per protocol condition and per held-out subject; baseline macro-F1 vs the macro-F1 after permuting each domain’s feature columns 20 times in the held-out subject’s test set; 95% CIs by 2,000 bootstraps over per-subject mean drops). Coordination metrics used per-subject, per (protocol condition × suit state) cell 10×10 Pearson correlation matrices of per-muscle 28-D feature trajectories, with mean |r|, mean r, and the network-density AUC over the threshold range (trapezoidal integration; step = 0.01). End-to-end latency was measured on Apple M1 across 5 cycles per protocol condition (N = 15 cycles total): event detection was timed once per file and divided by cycle count; per-cycle processing comprised preprocessing, phase indexing, feature extraction (F18 or F28), and inference. Inference-only latency used pre-computed features (5 warm-up + 1,000 measured calls). Environment identifiers were recorded (macOS 26.3 arm64; Python 3.11.5; numpy 1.26.4; scipy 1.15.2; scikit-learn 1.7.1; PyWavelets 1.8.0; commit 4e38a5b409e338c1e890bb64f86461fa4cd46926). RF inference used scikit-learn parallelism (n_jobs = − 1). The “~9 Hz throughput” mentioned in the body is the inverse of the F18 + RF mean end-to-end latency (≈ 1/0.115 s ≈ 8.7 Hz). Event-detection sensitivity varied σ ∈ {50, 100, 150} ms and min_interval ∈ {0.4, 0.5, 0.6} s; classification sensitivity used artificial phase-boundary shifts of ± 10/±20/±40 ms with F18 recomputed. Paired two-tailed t-tests with Benjamini–Hochberg correction (q < 0.05) were used for ablation analyses on subject-level (n = 10) within-subject paired metrics. Empirical null analyses were computed on the same balanced LOSO split manifest at the cycle-level endpoint. Dummy baselines used scikit-learn DummyClassifier outputs for most-frequent and stratified strategies, and a uniform-probability dummy evaluated through the same probability-aggregation and deterministic-argmax scoring pipeline, followed by the same cycle-level macro-F1 scoring and subject-level aggregation as the primary endpoint. Permutation-null references were generated at cycle level by permuting true cycle labels within each held-out-subject × train-condition × test-condition cell while holding model predictions fixed, with 500 repetitions per cell; null summaries are reported in Supplementary Table S14. For the abstention analysis (Fig. 4 a), the confidence score was the cycle-level maximum class probability. Predictions with confidence below a threshold τ were abstained; coverage was defined as the fraction of non-abstained cycles. The error curve reports the PE←NW pre-specified asymmetric error rate among non-abstained off-diagonal transfer-cell predictions (PE←NW denotes the error in which a true NW cycle is classified as PE). The threshold τ was swept across the observed confidence range and the resulting curves were aggregated across held-out subjects. Supporting analyses NMF activation-pattern visualization, coordination-network metrics, feature-module ablation, abstention analysis, and event-detection sensitivity analyses were retained as supporting or robustness readouts. These analyses document descriptive signal structure and processing sensitivity; they do not define the primary endpoint and are not interpreted as mechanistic evidence of altered muscle synergies or closed-loop control behavior. The benchmark is intended to test whether a future method can exceed the fixed null-referenced LOSO transfer boundary under the same manifest and scoring endpoint, not to establish that the current RF baseline is an adequate recognizer. Use of AI tools ChatGPT (OpenAI) was used only for non-substantive language polishing and literature-search query formulation. All references, claims, analyses, numerical results, figures, and tables were selected, generated, verified, and approved by the authors. No LLM was listed as an author. Results Dataset and primary LOSO benchmark From 10 healthy adult men, 29,993 EMG-derived gait cycles were retained after cycle-construction filters across three protocol-defined walking-condition bundles and three externally imposed suit states. The primary endpoint was cycle-level macro-F1 under LOSO, with one held-out subject excluded entirely from model fitting and one prediction produced per HS→next-HS cycle. Phase-level samples were used for model fitting, but the pre-specified operational endpoint was the cycle-level decision obtained by averaging retained phase-wise class probabilities within each cycle. Within the balanced LOSO manifest, the evaluation set was capped at ≤ 50 cycles per recording file across 90 recording files (3 protocol conditions × 3 suit states × 10 subjects), producing up to 4,500 cycle-level samples (≈ 1,500 per protocol condition) and 12,000 cycle×phase samples per protocol condition before NL10 filtering; after NL10 filtering, all 4,500 manifest-selected cycles contributed at least one retained phase probability and were used for cycle-level scoring; cell-wise retained counts are provided in Supplementary Table S3 . In the primary cycle-level endpoint, within-condition macro-F1 was 0.43 [0.34, 0.51] for level, 0.41 [0.33, 0.49] for slope, and 0.42 [0.34, 0.50] for stairs. Across protocol-condition transfers, macro-F1 ranged from 0.18 to 0.41. Thus, the main empirical finding is not high absolute recognizer performance but a clear subject-independent transfer boundary under protocol-condition changes. Protocol-condition transfer gaps The transfer matrix should be interpreted as a protocol-condition transfer matrix, not as a pure terrain-only generalization test. The weakest transfers were level treadmill → slope walkway and level treadmill → stair stepmill. Secondary cycle×phase readouts showed the same diagonal-versus-off-diagonal ordering and the same weak transfer cells, but they remain sensitivity analyses because the eight phase samples within a gait cycle are correlated. Because the primary model used phase-level training while the endpoint was cycle-level, direct cycle-level RF and multinomial logistic-regression comparators, trained and evaluated on one aggregated vector per cycle, were treated as hierarchy checks; RF reached 0.36–0.44 within condition and 0.16–0.44 across transfers, and multinomial logistic regression reached 0.37–0.45 within condition and 0.17–0.38 across transfers, while preserving the same diagonal-versus-off-diagonal ordering. Numerical comparator values are reported in Table 3 and Supplementary Table S13. See Fig. 2 for secondary cycle×phase transfer-cell ordering. Findings should be interpreted as protocol-condition transfer across bundled terrain/platform/speed conditions, not as terrain-only generalization. Table 3 Headline benchmark numbers. This condensed table keeps only the four headline messages for quick interpretation; detailed null-reference values and direct cycle-level comparator numerics are reported in Supplementary Tables S13–S14 and Additional file 8 under the same LOSO split manifest. Block Purpose Key finding Primary F28 cycle-level LOSO main benchmark within 0.41–0.43, transfer 0.18–0.41 Direct cycle-level comparator endpoint hierarchy check RF 0.36–0.44 (within), 0.16–0.44 (transfer); LR 0.37–0.45 (within), 0.17–0.38 (transfer); same ordering Calibration sensitivity secondary only weak cells improve only with known-condition reference Latency computational cost only F28 offline, F18 lower-latency comparator Analytical references for 3-class macro-F1 were approximately 0.333 for class-prior-matched random prediction and 0.168 for global-majority prediction. Under these references, the within-condition cells were modestly above null ranges, whereas several off-diagonal transfer cells were near null. Secondary calibration readouts under separate regimes Calibration-related results are reported as regime-specific secondary readouts rather than as deployment adaptation results. In secondary cycle×phase readouts, known-condition reference normalization increased macro-F1 in the two weakest transfer cells: level → slope by + 0.095 [0.04, 0.15] and level → stairs by + 0.098 [0.05, 0.15]. Because this regime requires the correct test protocol-condition label, these deltas are interpreted as oracle-style reference-normalization sensitivity rather than condition-agnostic adaptation or deployable online normalization. These values correspond to Table 3 (Block C) and Fig. 3 . The offline pooled subject-window normalization analysis provided an additional sensitivity readout using the first four unlabeled held-out-subject cycles per file/condition as an excluded calibration window. Because this window depended on recording-file structure, it is retained only as an offline reference analysis. The small-label LOTO-under-LOSO analysis characterized post-hoc calibration behavior under cycle-level label sampling and is not treated as an autonomous label-acquisition or deployment protocol. Class-wise confusion summaries in Additional file 3 indicate that the weakest transfer cells were characterized by concentrated off-diagonal confusion rather than uniform three-class degradation; therefore, the three-class results should be interpreted as protocol-state discriminability under the benchmark labels, not as isolated active-assistance detection. Computational latency comparator and supporting analyses In the offline replay latency benchmark on Apple M1, the F18 + RF path required 115 ms/cycle end-to-end, whereas the full F28 path required 2697 ms/cycle and remained an offline processed-feature reference. RF inference on pre-computed features required 1.36 ms on average. These numbers are computational readouts only and do not represent within-cycle closed-loop control latency. The latency analysis separates computational cost from recognition performance: the full processed-feature path is retained as the primary offline benchmark reference for discriminability profiling, whereas the lower-latency F18 path is reported as a computational comparator under the same split manifest; the two paths are not combined into a single recognizer or deployment claim. All latency measurements were obtained from an offline-replay, post-cycle computational benchmark on Apple M1; the cycle-level decision is available only after HS→next-HS cycle completion. Latency constraints and abstention summaries are linked in Table 3 (Block D) and Fig. 4 . Event-detection and phase-boundary sensitivity analyses did not define the primary endpoint but were used to check whether the main transfer ordering was dominated by plausible HS/TO proxy perturbations; full source values are provided in Additional file 8. Supporting NMF, coordination-network, feature-module ablation, abstention-policy, and event-sensitivity analyses are provided as descriptive or robustness readouts. They were not used to define the primary endpoint and are not interpreted as mechanistic evidence of altered muscle synergies, safety validation, or closed-loop control behavior. Discussion This study contributes a reusable null-referenced transfer-boundary benchmark for EMG-only recognition modules in wearable gait-assistance research. Its value lies not in high absolute recognition accuracy, but in defining where subject-independent transfer fails under fixed, processed-feature-level reproducible LOSO scoring. The modest macro-F1 values therefore provide a reproducible reliability boundary that future recognition methods, sensor-fusion pipelines, or calibration strategies should exceed under the same split manifest and scoring procedure. The contribution is a fixed benchmark procedure—split manifest, transfer grid, null references, scoring endpoint, and reproducibility package—rather than a new classifier or deployment controller. Calibration analyses are regime-specific secondary readouts. Known-condition reference normalization increased macro-F1 in selected weak transfer cells but required the correct test protocol-condition label. The pooled subject-window analysis remained an offline sensitivity readout because its calibration window depended on recording-file structure and was excluded from evaluation. The small-label LOTO-under-LOSO analysis characterized post-hoc behavior under cycle-level label sampling and is not a deployment policy. Because the ΦB ON-versus-OFF comparator did not show a uniform macro-F1 gain, ΦB should be interpreted only as a predefined label-independent pre-fusion scaling convention. The reported benchmark should therefore not be read as evidence that ΦB improves subject-independent recognition performance or as evidence for a mechanistic muscle-force model. The latency analysis should be read as an offline computational constraint analysis. F28 defines the higher-cost processed-feature reference used for the primary discriminability benchmark, whereas F18 defines a lower-latency computational comparator. Because the cycle-level decision is available only after HS→next-HS cycle completion, and because closed-loop control was not implemented, neither latency pathway demonstrates within-cycle control readiness. F18 performance analyses are retained as supplementary comparator readouts (Supplementary Table S10 ), while primary endpoint interpretation remains anchored to F28-based cycle-level LOSO evaluation. Supporting NMF, coordination-network, feature-module ablation, abstention-policy, and event-sensitivity analyses document descriptive signal structure and robustness checks. They are not presented as mechanistic evidence of altered muscle synergies, safety validation, or controller-readiness evidence. This separation is essential because the primary benchmark is a processed-feature recognition-stage evaluation. Limitations. The sample included only 10 healthy adult men and one cable-driven suit platform. Terrain, platform, and locomotor speed were jointly varied, so the transfer matrix is not a pure terrain-only test. EMG-derived event proxies were not validated against external sensors. The released benchmark is processed-feature based; raw EMG redistribution remains restricted under the stated DUA procedure. The protocol metadata did not explicitly log randomization/counterbalancing order, scheduled rest intervals, or electrode reattachment events, leaving residual order, fatigue, session, or electrode-state effects as interpretation boundaries. Finally, closed-loop control was not implemented, so no deployment-level control claim follows from these results. In addition, the anchor and cable-path descriptors define the intended assistance configuration only; body-surface anchor motion, strap-interface pressure, cable-line migration, and subject-specific delivered joint moments were not measured in this benchmark. Because phase-level samples from the same cycle are correlated, primary inference in the present report is based on subject-level uncertainty, and a cycle-weighted RF sensitivity was additionally reported as a robustness check (Additional file 8). Direct cycle-level aggregation comparators and empirical dummy/permutation null baselines were added in this revision as robustness analyses on the same LOSO split (Supplementary Tables S13-S14, Additional file 8). These analyses reduce, but do not eliminate, interpretation uncertainty around phase-level training versus cycle-level endpoint hierarchy. Because all participants were healthy adult men, sex- or gender-related variability could not be evaluated, and these findings should not be generalized to women, older adults, or clinical populations without external validation. Importantly, the released package supports reproducibility of the processed-feature recognition benchmark, but does not fully reproduce actuator-level assistance waveforms; therefore PE-related findings should be interpreted as protocol-state discriminability under a restricted assistance descriptor rather than as evidence of a specific torque-profile effect. Conclusions This proof-of-concept study provides a leakage-controlled processed-feature benchmark for sEMG-based discriminability of externally imposed NW/UE/PE protocol states in a cable-driven walking-assistance suit. Under LOSO evaluation, cycle-level macro-F1 was modest within condition and degraded under protocol-condition transfer, demonstrating substantial subject-independent transfer gaps. The benchmark therefore defines a reproducible transfer-boundary and method-comparison reference rather than a deployment-ready recognition or control system. External validation across larger and more diverse populations, alternative suit platforms, decoupled speed protocols, externally validated gait events, and closed-loop experiments is required before deployment-level claims can be made. Abbreviations ApEn approximate entropy BF biceps femoris CI confidence interval CRediT Contributor Roles Taxonomy cycle×phase gait cycle by fixed—percentage cycle bin based on Perry timing references db5 Daubechies—5 wavelet DUA Data Use Agreement EMG electromyography F18 18—feature time/frequency/wavelet vector F28 F18 + NL10 FD fractal dimension FT fast—fibre fraction GM gastrocnemius medialis HS heel strike IC initial contact IITP Institute of Information & Communications Technology Planning & Evaluation IRB Institutional Review Board ISw initial swing JNER Journal of NeuroEngineering and Rehabilitation LLM large language model LOSO leave—one—subject—out LOTO leave—one—protocol—condition—out LR loading response MAV mean absolute value MAD median absolute deviation MSt mid stance MSw mid swing MSIT Ministry of Science and ICT NL10 10—feature nonlinear—dynamics vector NW not worn PCSA physiological cross—sectional area PE worn—active suit PSw pre—swing RF Random Forest (classifier) or rectus femoris (muscle, by context) RMS root mean square RQA recurrence quantification analysis SampEn sample entropy SD standard deviation SENIAM Surface ElectroMyoGraphy for the Non—Invasive Assessment of Muscles sEMG surface electromyography SSC slope sign change TA tibialis anterior TFL tensor fasciae latae TO toe—off TSt terminal stance TSw terminal swing UE worn—inactive suit WEF wavelet ensemble fusion WL waveform length ZCR zero—crossing rate. Declarations Trial registration Not applicable. This study is a recognition-stage methodological benchmark and does not evaluate prospectively assigned health-care interventions against clinical health outcomes. Ethics approval and consent to participate This study was approved by the Institutional Review Board (IRB) of the Electronics and Telecommunications Research Institute (ETRI) under approval number N01-202409-01-020. All participants were healthy adult men aged ≥19 years and provided written informed consent prior to enrollment, including consent for the experimental procedures described in Methods. All procedures were performed in accordance with the Declaration of Helsinki and the relevant guidelines and regulations. Consent for publication Not applicable. This manuscript contains no individually identifiable participant data, images, or videos; all figures and tables present only de-identified, aggregate, or schematic content. Availability of data and materials The openly reproducible component of this study is the processed-feature recognition benchmark. The Zenodo deposit [44] contains processed feature tensors, metadata, the official split manifest (splits/loso_manifest_v1.csv), official scoring script (metrics/score.py), H-MD-WEF + RF baseline source code, source values, reproduce_all.sh, and the reproducibility lock (seed = 42; macOS 26.3 arm64; Python 3.11.5; numpy 1.26.4; scipy 1.15.2; scikit-learn 1.7.1; commit 4e38a5b409e338c1e890bb64f86461fa4cd46926), which reproduce the reported Tables/Figures at the processed-feature level. Raw-signal preprocessing and EMG-derived HS/TO event extraction are documented for transparency but are not fully rerunnable from the public package because raw EMG redistribution is restricted by IRB-approved consent. During peer review, an anonymous reviewer-only Zenodo access link is provided through the JNER editorial system and converted to the public DOI on acceptance. Raw EMG may be requested for non-commercial academic research under a Data Use Agreement, subject to IRB/consent constraints. Competing interests The authors declare no competing interests. Funding This work was supported by the Institute of Information & Communications Technology Planning & Evaluation (IITP), funded by the Ministry of Science and ICT (MSIT), Republic of Korea (Project No. RS-2022-II220025, “Development of soft-suit technology to support human motor ability”). The funder had no role in study design; data collection, analysis, or interpretation; manuscript preparation; or the decision to publish. Authors’ contributions CRediT contributor roles (initials): SBL (Song-Bi Lee): Conceptualization; Methodology; Formal analysis; Data curation; Visualization; Writing – original draft. DWL (Dong-Woo Lee): Data curation; Funding acquisition; Project administration; Writing – review & editing. YK (Yongjun Kim): Data curation; Software; Validation; Writing – review & editing. GH (Gisu Heo): Writing – review & editing. CO (Changmok Oh): Writing – review & editing. SE (Suyeong Eom): Data curation; IRB protocol development. All authors read and approved the final manuscript. Acknowledgements We thank all participants. We used Elsevier Language Editing Services for English-language editing of this manuscript. Use of large language models See Methods: Use of AI tools. Reporting guideline checklist A completed reporting guideline checklist (STROBE, with item-to-manuscript mapping and N/A justifications where appropriate) is included as a separate submission file for peer review. Because the present study is a machine-learning prediction-style recognition benchmark rather than a clinical observational study, we additionally provide a TRIPOD+AI-style item-mapping supplement that addresses model reporting items that are not fully covered by STROBE (data split definition, leakage prevention, calibration regimes, missing/filtered samples, model hyperparameters, performance uncertainty, and the absence of external validation). The TRIPOD+AI mapping is intended only as supplementary transparency, not as a claim that this work is a clinical prediction model. References Sawicki, G. S. et al. The exoskeleton expansion: improving walking and running economy. J. NeuroEngineering Rehabil. 17, 25 (2020). Young, A. J. & Ferris, D. P. State of the art and future directions for lower limb robotic exoskeletons. IEEE Trans. Neural Syst. Rehabil. Eng. 25, 171–182 (2017). Panizzolo, F. A. et al. A biologically-inspired multi-joint soft exosuit that can reduce the energy cost of loaded walking. J. NeuroEngineering Rehabil. 13, 43 (2016). Ding, Y. et al. Human-in-the-loop optimization of hip assistance with a soft exosuit during walking. Sci. Robot. 3, eaar5438 (2018). Zhang, J. et al. Human-in-the-loop optimization of exoskeleton assistance during walking. Science 356, 1280–1284 (2017). Malcolm, P. et al. A simple exoskeleton that assists plantarflexion can reduce the metabolic cost of human walking. PLoS ONE 8, e56137 (2013). Quinlivan, B. T. et al. Assistance magnitude versus metabolic cost reductions for a tethered multiarticular soft exosuit. Sci. Robot. 2, eaah4416 (2017). Huang, H. et al. Continuous locomotion-mode identification for prosthetic legs based on neuromuscular-mechanical fusion. IEEE Trans. Biomed. Eng. 58, 2867–2875 (2011). Phinyomark, A. et al. Feature reduction and selection for EMG signal classification. Expert Syst. Appl. 39, 7420–7431 (2012). Englehart, K. & Hudgins, B. A robust, real-time control scheme for multifunction myoelectric control. IEEE Trans. Biomed. Eng. 50, 848–854 (2003). Scheme, E. & Englehart, K. Electromyogram pattern recognition for control of powered upper-limb prostheses: state of the art and challenges for clinical use. J. Rehabil. Res. Dev. 48, 643–659 (2011). Farina, D. et al. The extraction of neural information from the surface EMG for the control of upper-limb prostheses: emerging avenues and challenges. IEEE Trans. Neural Syst. Rehabil. Eng. 22, 797–809 (2014). Hargrove, L. J. et al. Myoelectric pattern recognition outperforms direct control for transhumeral amputees with targeted muscle reinnervation: a randomized clinical trial. Sci. Rep. 7, 13840 (2017). Liu, J. et al. EMG-based real-time linear-nonlinear cascade regression decoding of shoulder, elbow and wrist movements in able-bodied persons and stroke survivors. IEEE Trans. Biomed. Eng. 67, 1272–1281 (2020). Nazarpour, K. et al. A note on the probability distribution function of the surface electromyogram signal. Brain Res. Bull. 90, 73–79 (2013). De Luca, C. J. The use of surface electromyography in biomechanics. J. Appl. Biomech. 13, 135–163 (1997). Merletti, R. & Farina, D. Analysis of intramuscular electromyogram signals. Phil. Trans. R. Soc. A 367, 357–368 (2009). Clancy, E. A. et al. Sampling, noise-reduction and amplitude estimation issues in surface electromyography. J. Electromyogr. Kinesiol. 12, 1–16 (2002). Phinyomark, A. et al. EMG feature evaluation for improving myoelectric pattern recognition robustness. Expert Syst. Appl. 40, 4832–4840 (2013). Tkach, D. et al. Study of stability of time-domain features for electromyographic pattern recognition. J. NeuroEngineering Rehabil. 7, 21 (2010). Hudgins, B. et al. A new strategy for multifunction myoelectric control. IEEE Trans. Biomed. Eng. 40, 82–94 (1993). Graupe, D. & Cline, W. K. Functional separation of EMG signals via ARMA identification methods for prosthesis control purposes. IEEE Trans. Syst. Man Cybern. 5, 252–259 (1975). Chan, A. D. C. & Englehart, K. B. Continuous myoelectric control for powered prostheses using hidden Markov models. IEEE Trans. Biomed. Eng. 52, 121–124 (2005). Mallat, S. G. A theory for multiresolution signal decomposition: the wavelet representation. IEEE Trans. Pattern Anal. Mach. Intell. 11, 674–693 (1989). Daubechies, I. Ten Lectures on Wavelets (SIAM, 1992). Englehart, K. et al. A wavelet-based continuous classification scheme for multifunction myoelectric control. IEEE Trans. Biomed. Eng. 48, 302–311 (2001). Richman, J. S. & Moorman, J. R. Physiological time-series analysis using approximate entropy and sample entropy. Am. J. Physiol. Heart Circ. Physiol. 278, H2039–H2049 (2000). Higuchi, T. Approach to an irregular time series on the basis of the fractal theory. Physica D 31, 277–283 (1988). Hausdorff, J. M. Gait dynamics in Parkinson's disease: common and distinct behavior among stride length, gait variability, and fractal-like scaling. Chaos 19, 026113 (2009). Hausdorff, J. M. et al. Gait variability and fall risk in community-living older adults: a 1-year prospective study. Arch. Phys. Med. Rehabil. 82, 1050–1056 (2001). Rosenstein, M. T., Collins, J. J. & De Luca, C. J. A practical method for calculating largest Lyapunov exponents from small data sets. Physica D 65, 117–134 (1993). Katz, M. J. Fractals and the analysis of waveforms. Comput. Biol. Med. 18, 145–156 (1988). Pincus, S. M. Approximate entropy as a measure of system complexity. Proc. Natl. Acad. Sci. USA 88, 2297–2301 (1991). Bandt, C. & Pompe, B. Permutation entropy: a natural complexity measure for time series. Phys. Rev. Lett. 88, 174102 (2002). Hurst, H. E. Long-term storage capacity of reservoirs. Trans. Am. Soc. Civ. Eng. 116, 770–799 (1951). Webber, C. L. Jr. & Zbilut, J. P. Dynamical assessment of physiological systems and states using recurrence plot strategies. J. Appl. Physiol. 76, 965–973 (1994). Marwan, N. et al. Recurrence plots for the analysis of complex systems. Phys. Rep. 438, 237–329 (2007). Hermens, H. J. et al. Development of recommendations for SEMG sensors and sensor placement procedures. J. Electromyogr. Kinesiol. 10, 361–374 (2000). Perry, J. & Burnfield, J. M. Gait Analysis: Normal and Pathological Function 2nd edn (SLACK, 2010). Ward, S. R. et al. Are current measurements of lower extremity muscle architecture accurate? Clin. Orthop. Relat. Res. 467, 1074–1082 (2009). Johnson, M. A., Polgar, J., Weightman, D. & Appleton, D. Data on the distribution of fibre types in thirty-six human muscles: an autopsy study. J. Neurol. Sci. 18, 111–129 (1973). Pedregosa, F. et al. Scikit-learn: machine learning in Python. J. Mach. Learn. Res. 12, 2825–2830 (2011). Breiman, L. Random forests. Mach. Learn. 45, 5–32 (2001). Lee, S.-B. et al. MD-WEF recognition-stage benchmark v1.0 (processed-feature tensors, split manifest, scoring scripts, and baseline code) [dataset]. Zenodo. https://doi.org/10.5281/zenodo.19547468. Additional Declarations No competing interests reported. Supplementary Files 07Additionalfile1.zip 08Additionalfile2.zip 09Additionalfile3.zip 10Additionalfile4.zip 11Additionalfile5.zip 12Additionalfile6.zip 13Additionalfile7.zip Cite Share Download PDF Status: Under Review Version 1 posted Reviewers agreed at journal 11 May, 2026 Reviewers agreed at journal 09 May, 2026 Reviewers agreed at journal 08 May, 2026 Reviewers agreed at journal 07 May, 2026 Reviewers invited by journal 04 May, 2026 Editor assigned by journal 04 May, 2026 Submission checks completed at journal 04 May, 2026 First submitted to journal 01 May, 2026 You are reading this latest preprint version Research Square lets you share your work early, gain feedback from the community, and start making changes to your manuscript prior to peer review in a journal. As a division of Research Square Company, we’re committed to making research communication faster, fairer, and more useful. We do this by developing innovative software and high quality services for the global research community. Our growing team is made up of researchers and industry professionals working together to solve the most critical problems facing scientific publishing. Also discoverable on Platform About Our Team In Review Editorial Policies Advisory Board Help Center Resources Author Services Accessibility API Access RSS feed Manage Cookie Preferences © Research Square 2026 | ISSN 2693-5015 (online) Privacy Policy Terms of Service Do Not Sell My Personal Information {"props":{"pageProps":{"initialData":{"identity":"rs-9590227","acceptedTermsAndConditions":true,"allowDirectSubmit":false,"archivedVersions":[],"articleType":"Research Article","associatedPublications":[],"authors":[{"id":638645494,"identity":"4af9270f-9caa-4b5c-b12f-42b89974c7f2","order_by":0,"name":"Song-Bi Lee¹","email":"data:image/png;base64,iVBORw0KGgoAAAANSUhEUgAAAZAAAAAyAQMAAABI0h/eAAAABlBMVEX///8AAABVwtN+AAAACXBIWXMAAA7EAAAOxAGVKw4bAAAA4klEQVRIiWNgGAWjYJCCA2CSmRlES8gQVM6D0MKWANLCQ5QWGNMATQAHsGc/e/DAjz82efLuPJ9f3aix4GFgP3x0A15bePISDva2pRUbHubdZp1zDOgwnrS0G/gdlmNwgLfhcOLGZt5txjlsQC0SPGb4tfC/MTj45w9IC88z45x/xGiRyDE4zMN2OHE+Mw/z49w2YrTceGNwWLYtLXEDM5sZc26fBA8bIb+w9+cYf3zzxyZxfv/hx59zvtXJ8bMfPoZXCxwYHGBgkwAx2IhSDgLyDQzMH4hWPQpGwSgYBSMKAABSckaaDHFmQgAAAABJRU5ErkJggg==","orcid":"","institution":"Electronics and Telecommunications Research Institute (ETRI)","correspondingAuthor":true,"prefix":"","firstName":"Song-Bi","middleName":"","lastName":"Lee¹","suffix":""},{"id":638645495,"identity":"a34c1e79-ceaf-471c-820a-78ef87108e33","order_by":1,"name":"Yongjun Kim¹","email":"","orcid":"","institution":"Electronics and Telecommunications Research Institute (ETRI)","correspondingAuthor":false,"prefix":"","firstName":"Yongjun","middleName":"","lastName":"Kim¹","suffix":""},{"id":638645498,"identity":"64a38c14-d65a-44ea-9991-733f4ffc6b27","order_by":2,"name":"Gisu Heo¹","email":"","orcid":"","institution":"Electronics and Telecommunications Research Institute (ETRI)","correspondingAuthor":false,"prefix":"","firstName":"Gisu","middleName":"","lastName":"Heo¹","suffix":""},{"id":638645502,"identity":"ed86ab45-8624-45a5-b2e0-953d3fd00ce0","order_by":3,"name":"Changmok Oh¹","email":"","orcid":"","institution":"Electronics and Telecommunications Research Institute (ETRI)","correspondingAuthor":false,"prefix":"","firstName":"Changmok","middleName":"","lastName":"Oh¹","suffix":""},{"id":638645503,"identity":"f374fae2-37ff-40fa-a7a2-ceb90ec52134","order_by":4,"name":"Suyeong Eom¹","email":"","orcid":"","institution":"Electronics and Telecommunications Research Institute (ETRI)","correspondingAuthor":false,"prefix":"","firstName":"Suyeong","middleName":"","lastName":"Eom¹","suffix":""},{"id":638645504,"identity":"c3bde347-1726-450e-ba6c-e176fad356b2","order_by":5,"name":"Dong-Woo Lee¹","email":"","orcid":"","institution":"Electronics and Telecommunications Research Institute (ETRI)","correspondingAuthor":false,"prefix":"","firstName":"Dong-Woo","middleName":"","lastName":"Lee¹","suffix":""}],"badges":[],"createdAt":"2026-05-02 03:08:16","currentVersionCode":1,"declarations":"","doi":"10.21203/rs.3.rs-9590227/v1","doiUrl":"https://doi.org/10.21203/rs.3.rs-9590227/v1","draftVersion":[],"editorialEvents":[],"editorialNote":"","failedWorkflow":false,"files":[{"id":109175967,"identity":"2510b36c-30aa-4785-819e-ba73f90471f4","added_by":"auto","created_at":"2026-05-13 09:29:38","extension":"jpg","order_by":1,"title":"Figure 1","display":"","copyAsset":false,"role":"figure","size":941712,"visible":true,"origin":"","legend":"\u003cp\u003eStudy overview and recognition-stage benchmark pipeline.\u003c/p\u003e\n\u003cp\u003e(a) Computational pipeline schematic (workflow and endpoint hierarchy): signal acquisition and gait segmentation (EMG-derived HS/TO proxies, not externally validated gait events, with fixed-percentage cycle bins based on Perry timing references), feature extraction and fusion, normalization, Random Forest classification, phase-wise outputs, and cycle-level primary-endpoint aggregation, with a post-hoc calibration branch. (b) Benchmark evaluation-task grouping aligned with the Table 2 regimes: generalization (Tasks 1–2), calibration (Tasks 3–5), the latency benchmark, and the asymmetric-error readout. Regime summaries are provided in Table 2; full definitions and the supplementary index mapping are provided in Additional file 1. The phase labels denote fixed-percentage cycle bins derived from EMG-based cycle segmentation and should not be interpreted as externally validated biomechanical gait-phase annotations.\u003c/p\u003e","description":"","filename":"fig1.jpg","url":"https://assets-eu.researchsquare.com/files/rs-9590227/v1/61453f09806a6ed1b4382c98.jpg"},{"id":109175959,"identity":"c21e4f9c-7d35-4897-9386-0aa984756cf9","added_by":"auto","created_at":"2026-05-13 09:29:35","extension":"jpg","order_by":2,"title":"Figure 2","display":"","copyAsset":false,"role":"figure","size":348918,"visible":true,"origin":"","legend":"\u003cp\u003ePrimary endpoint matrix with secondary transfer readouts across protocol-defined walking-condition bundles.\u003c/p\u003e\n\u003cp\u003ePanel (a) shows the primary endpoint (cycle-level macro-F1, identical to the 3 × 3 matrix in Table 1b). Panels (b–d) are secondary cycle×phase sensitivity readouts: (b) baseline transfer matrix, (c) Δ heatmap defined as (known-condition reference normalization − baseline), and (d) off-diagonal/diagonal ratio summary (each cell normalized by the diagonal of its test-condition column). Panels (a) and (b) share the same transfer-cell layout; the color scale of panel (b) is shown explicitly to avoid over-reading absolute magnitude differences relative to panel (a). Known-condition reference normalization assumes that the test-time protocol-defined condition label is known and is therefore not a fully condition-agnostic unsupervised deployment setting.\u003c/p\u003e","description":"","filename":"fig2.jpg","url":"https://assets-eu.researchsquare.com/files/rs-9590227/v1/bee2f7460c6f21b65da326f3.jpg"},{"id":109175960,"identity":"e833c031-e531-4ccc-b643-bb4a25bebeed","added_by":"auto","created_at":"2026-05-13 09:29:35","extension":"jpg","order_by":3,"title":"Figure 3","display":"","copyAsset":false,"role":"figure","size":359182,"visible":true,"origin":"","legend":"\u003cp\u003eSecondary calibration and PE←NW asymmetric-error readouts.\u003c/p\u003e\n\u003cp\u003ePanels in this figure are not all directly comparable because they mix different endpoint units and calibration regimes. (a) Regime-specific calibration readout for key transfer cells (level→slope and level→stairs) across baseline, known-condition reference normalization requiring the test protocol-condition label but no held-out suit-state labels, and offline pooled subject-window reference normalization. (b) Small-label post-hoc calibration — a separate LOTO-under-LOSO regime, not the primary endpoint — plotted against cycle-level sampled label budgets reported as phase-label-equivalent units. (c) PE←NW pre-specified asymmetric error rates within each calibration regime; numbers should be compared only within the same regime. This panel is an offline asymmetric error readout only and is not a safety endpoint. All calibration panels are regime-specific offline readouts and should not be interpreted as a validated online adaptation or safety procedure.\u003c/p\u003e","description":"","filename":"fig3.jpg","url":"https://assets-eu.researchsquare.com/files/rs-9590227/v1/4d0ecc5749dd8cf17ac37a2e.jpg"},{"id":109176012,"identity":"20a9227a-6597-480d-8e9b-095bc2bedde3","added_by":"auto","created_at":"2026-05-13 09:30:14","extension":"jpg","order_by":4,"title":"Figure 4","display":"","copyAsset":false,"role":"figure","size":329341,"visible":true,"origin":"","legend":"\u003cp\u003eAbstain-policy and latency benchmark summary.\u003c/p\u003e\n\u003cp\u003e(a) Probability-based abstention policy on off-diagonal transfer cells, showing the coverage–macro-F1 and coverage–PE←NW asymmetric-error trade-off curves. PE←NW denotes the pre-specified asymmetric error in which a true NW cycle is classified as PE. (b) RF inference only on pre-computed feature vectors, compared with a 10-ms computational reference interval (the 10-ms reference is a computational marker only and not evidence of online closed-loop readiness). (c) End-to-end per-cycle latency by pathway, shown as empirical mean, p95, and p99 values. Component stages were timed to obtain the total latency but are not plotted separately. All values come from an offline-replay, post-cycle computational benchmark (N = 15 cycles total) and should not be interpreted as online within-cycle closed-loop control latencies. The F18 path is reported as a lower-latency computational comparator on the same split manifest, not as an online closed-loop control claim.\u003c/p\u003e","description":"","filename":"fig4.jpg","url":"https://assets-eu.researchsquare.com/files/rs-9590227/v1/53bc53c25d395ae5acebf60f.jpg"},{"id":109206691,"identity":"84edd284-5b3b-421e-87c8-7d8ef2fb661b","added_by":"auto","created_at":"2026-05-13 15:15:22","extension":"pdf","order_by":0,"title":"","display":"","copyAsset":false,"role":"manuscript-pdf","size":2286635,"visible":true,"origin":"","legend":"","description":"","filename":"manuscript.pdf","url":"https://assets-eu.researchsquare.com/files/rs-9590227/v1/8b084eda-c2b2-4d8b-b4d8-37c5eee45298.pdf"},{"id":109176011,"identity":"e0a1c2b8-4df7-491a-b8a9-77debdfe3ce4","added_by":"auto","created_at":"2026-05-13 09:30:08","extension":"zip","order_by":1,"title":"","display":"","copyAsset":false,"role":"supplement","size":38616,"visible":true,"origin":"","legend":"","description":"","filename":"07Additionalfile1.zip","url":"https://assets-eu.researchsquare.com/files/rs-9590227/v1/13792029840f6489178c510e.zip"},{"id":109175996,"identity":"2a675f30-8a4b-424a-bf69-29129ad0fc4f","added_by":"auto","created_at":"2026-05-13 09:29:58","extension":"zip","order_by":2,"title":"","display":"","copyAsset":false,"role":"supplement","size":98835,"visible":true,"origin":"","legend":"","description":"","filename":"08Additionalfile2.zip","url":"https://assets-eu.researchsquare.com/files/rs-9590227/v1/5bc8f34971d9c4c9794d6d26.zip"},{"id":109175968,"identity":"ecda5675-5203-417d-bf30-6c0cda57d641","added_by":"auto","created_at":"2026-05-13 09:29:47","extension":"zip","order_by":3,"title":"","display":"","copyAsset":false,"role":"supplement","size":266821,"visible":true,"origin":"","legend":"","description":"","filename":"09Additionalfile3.zip","url":"https://assets-eu.researchsquare.com/files/rs-9590227/v1/74696c8aa879c56b01363585.zip"},{"id":109176008,"identity":"286a727d-b84e-4d19-9d60-84e128952abc","added_by":"auto","created_at":"2026-05-13 09:30:07","extension":"zip","order_by":4,"title":"","display":"","copyAsset":false,"role":"supplement","size":385660,"visible":true,"origin":"","legend":"","description":"","filename":"10Additionalfile4.zip","url":"https://assets-eu.researchsquare.com/files/rs-9590227/v1/774112fcb0cc78595ca2e20b.zip"},{"id":109175961,"identity":"61b0e07d-169f-4abb-9db2-d6e71e63819c","added_by":"auto","created_at":"2026-05-13 09:29:35","extension":"zip","order_by":5,"title":"","display":"","copyAsset":false,"role":"supplement","size":163006,"visible":true,"origin":"","legend":"","description":"","filename":"11Additionalfile5.zip","url":"https://assets-eu.researchsquare.com/files/rs-9590227/v1/860ad21fa6c6ad998e6e6875.zip"},{"id":109176014,"identity":"8fa38f10-2a8c-4828-918a-3c5e744e4fe6","added_by":"auto","created_at":"2026-05-13 09:30:14","extension":"zip","order_by":6,"title":"","display":"","copyAsset":false,"role":"supplement","size":93604,"visible":true,"origin":"","legend":"","description":"","filename":"12Additionalfile6.zip","url":"https://assets-eu.researchsquare.com/files/rs-9590227/v1/f31556a9d976c7451d645fd1.zip"},{"id":109205547,"identity":"1a8ba003-36d5-45e9-8598-6d31f7f15718","added_by":"auto","created_at":"2026-05-13 15:05:40","extension":"zip","order_by":7,"title":"","display":"","copyAsset":false,"role":"supplement","size":7541,"visible":true,"origin":"","legend":"","description":"","filename":"13Additionalfile7.zip","url":"https://assets-eu.researchsquare.com/files/rs-9590227/v1/bbcb23e66b3145b43764f2ae.zip"}],"financialInterests":"No competing interests reported.","formattedTitle":"A reproducible processed-feature LOSO benchmark for sEMG protocol-state discriminability in a cable-driven walking-assistance suit","fulltext":[{"header":"Background","content":"\u003cp\u003eWearable gait-assistance systems such as cable-driven suits may require real-time discrimination of user assistance states for adaptive closed-loop control [\u003cspan additionalcitationids=\"CR2 CR3 CR4 CR5 CR6\" citationid=\"CR1\" class=\"CitationRef\"\u003e1\u003c/span\u003e\u0026ndash;\u003cspan citationid=\"CR7\" class=\"CitationRef\"\u003e7\u003c/span\u003e]. Surface electromyography (sEMG) can non-invasively measure muscle activity and has been used for gait-mode identification, prosthetic control, and classification [\u003cspan additionalcitationids=\"CR9 CR10 CR11 CR12 CR13\" citationid=\"CR8\" class=\"CitationRef\"\u003e8\u003c/span\u003e\u0026ndash;\u003cspan citationid=\"CR14\" class=\"CitationRef\"\u003e14\u003c/span\u003e]. A wide variety of time-domain, frequency-domain, and wavelet features have been proposed for EMG classification [\u003cspan citationid=\"CR9\" class=\"CitationRef\"\u003e9\u003c/span\u003e, \u003cspan additionalcitationids=\"CR16 CR17 CR18 CR19 CR20 CR21 CR22 CR23 CR24 CR25\" citationid=\"CR15\" class=\"CitationRef\"\u003e15\u003c/span\u003e\u0026ndash;\u003cspan citationid=\"CR26\" class=\"CitationRef\"\u003e26\u003c/span\u003e], and nonlinear dynamics measures can capture signal complexity associated with gait variability [\u003cspan additionalcitationids=\"CR28 CR29 CR30 CR31 CR32 CR33 CR34 CR35 CR36\" citationid=\"CR27\" class=\"CitationRef\"\u003e27\u003c/span\u003e\u0026ndash;\u003cspan citationid=\"CR37\" class=\"CitationRef\"\u003e37\u003c/span\u003e]. However, the multi-terrain assistance-state sEMG classification field has not converged on a shared definition of evaluation units, allowed test-time information, and risk/latency metrics, so a single number \u0026mdash; classifier accuracy\u0026thinsp;=\u0026thinsp;X% \u0026mdash; is rarely sufficient to compare two systems under matched conditions.\u003c/p\u003e \u003cp\u003eDespite this rich methodological landscape, subject-independent sEMG-based protocol-state discriminability still lacks an agreed evaluation grid that specifies which subjects, which terrain/platform/speed bundle conditions, and which test-time information are allowed for comparison. This proof-of-concept study defines and evaluates a leakage-controlled recognition-stage benchmark for sEMG-based discriminability of externally imposed NW/UE/PE protocol states in a single cable-driven suit platform. The primary objective is to quantify subject-independent cycle-level macro-F1 under within-condition LOSO and across protocol-defined terrain/platform/speed bundle transfer cells. Secondary objectives are to document regime-specific calibration behavior and computational latency constraints, without treating these analyses as deployment-readiness evidence. To support reproducible comparison, we release a fixed split manifest, scoring script, fixed seeds, environment lock, and processed-feature package. Thus, the methodological contribution is a bounded benchmark protocol rather than a new classifier architecture: it specifies the evaluation grid, allowed test-time information, null references, calibration regimes, and cycle-level scoring endpoint needed for reproducible comparison of future recognition methods.\u003c/p\u003e \u003cp\u003eOperational rationale of the benchmark label. The purpose is not to infer a suit controller state that the system already knows, but to quantify whether wearable-hardware presence and active-assistance protocol leave measurable subject-independent signatures in the sEMG feature space under leakage-controlled evaluation. Accordingly, NW/UE/PE are used as protocol-state discriminability labels for benchmark comparison, not as a substitute for controller-state telemetry.\u003c/p\u003e \u003cp\u003eRelevance to JNER readership. Quantifying when subject-independent transfer fails is practically important for neuroengineering and rehabilitation-assistance pipelines because it defines the reliability boundary of EMG-only recognition modules before they are coupled to adaptive control or clinical translation workflows.\u003c/p\u003e"},{"header":"Methods","content":"\u003cp\u003ePhase-training and cycle-level endpoint boundary. The model was trained on phase-bin samples, but all primary performance claims use one prediction per gait cycle. Because phase bins within a cycle are correlated, uncertainty was estimated only at the held-out-subject level. Direct cycle-level RF and logistic-regression comparators were included to check that the primary transfer ordering was not an artifact of phase-level training.\u003c/p\u003e \u003cdiv id=\"Sec3\" class=\"Section2\"\u003e \u003ch2\u003eParticipants\u003c/h2\u003e \u003cp\u003eTen healthy adult men (age 25.3\u0026thinsp;\u0026plusmn;\u0026thinsp;3.2 years; height 168.4\u0026thinsp;\u0026plusmn;\u0026thinsp;8.7 cm; weight 64.2\u0026thinsp;\u0026plusmn;\u0026thinsp;11.3 kg) participated. The study was approved by the Institutional Review Board (IRB) of the Electronics and Telecommunications Research Institute (ETRI) under approval number N01-202409-01-020, and written informed consent was obtained from every participant. No a priori sample-size calculation was performed because this was a proof-of-concept benchmark based on the available cable-driven suit dataset; all eligible participants with complete NW/UE/PE recordings across the three protocol-defined walking conditions were included.\u003c/p\u003e \u003c/div\u003e\n\u003ch3\u003eCable-driven suit and suit states\u003c/h3\u003e\n\u003cp\u003eIn this benchmark, NW, UE, and PE are externally imposed protocol-state labels. NW denotes walking without the suit, UE denotes the suit worn with assistance disabled, and PE denotes the same worn configuration with active assistance enabled under a fixed gait-synchronous study controller. UE and PE therefore differ at the protocol level by actuator active/inactive status, not by suit placement or wearable configuration. The classifier is not intended to infer a controller state unknown to the device, estimate torque magnitude, reconstruct actuator commands, or decode user intent. Instead, the labels are used to test whether wearable-hardware presence and active-assistance protocol leave subject-independent signatures in the processed sEMG feature space.\u003c/p\u003e \u003cp\u003eThe released benchmark package supports reproduction of the recognition-stage endpoint, not reproduction of the actuator-control waveform. Controller-command traces, continuous force/torque time series, internal controller states, update-rate details, and trigger thresholds are excluded from the released predictors and scoring inputs. Minimum non-identifying protocol metadata are provided to define the PE-versus-UE contrast at the benchmark level.\u003c/p\u003e \u003cp\u003eMinimum PE descriptor and release boundary. UE and PE used the same wearable hardware configuration: a cable-driven soft suit with a back-mounted actuator unit and strap-fixed thigh/pelvic anchor interfaces. In PE, four motor-driven cable channels provided gait-synchronous bilateral hip flexion/extension-related assistance through the anchor path; in UE, assistance was disabled under the same worn configuration and fixed study-controller settings. No subject-specific controller retuning or hip-joint-center reconstruction was used. The released predictors include only processed sEMG features and protocol metadata; actuator commands, cable-tension/force\u0026ndash;torque traces, and internal controller states were excluded. Therefore, PE is interpreted as a categorical active-assistance protocol state, and detailed release-boundary descriptors are provided in Supplementary Table S0 and Additional file 1.\u003c/p\u003e \u003cp\u003eMinimum non-identifying PE protocol metadata and the benchmark release boundary are summarized in Supplementary Table S0.\u003c/p\u003e\n\u003ch3\u003eWalking protocol and protocol-condition definition\u003c/h3\u003e\n\u003cp\u003eData were collected under three suit states (NW, UE, PE) and three protocol-defined walking conditions. Level walking was performed on a treadmill at 0.83 m/s. Stair walking was performed on an Egojin Highclimb stepmill (step height 21 cm) at 0.54 m/s. Slope walking was performed on a 25 m walkway at an 18\u0026deg; gradient at self-selected speed (50 m round trip; trial-mean speed 0.227\u0026thinsp;\u0026plusmn;\u0026thinsp;0.016 m/s). Mean recording durations were 299\u0026thinsp;\u0026plusmn;\u0026thinsp;23 s (level), 221\u0026thinsp;\u0026plusmn;\u0026thinsp;17 s (slope), and 180\u0026thinsp;\u0026plusmn;\u0026thinsp;6 s (stairs). Each 18\u0026deg; slope round trip included two opposite-direction segments at the same gradient, which were treated as uphill and downhill in analysis. Uphill and downhill segments were retained within the single slope protocol-condition bundle for the primary benchmark rather than analyzed as separate primary conditions. Protocol metadata included subject, protocol condition, suit state, and trial/run identifiers; randomization/counterbalancing order, scheduled rest intervals, and electrode reattachment logs were not explicitly recorded in the released benchmark metadata. Protocol-condition metadata are summarized in Additional file 1 Table A2.\u003c/p\u003e \u003cp\u003eThroughout this paper, \u0026ldquo;protocol condition\u0026rdquo; denotes the (terrain\u0026thinsp;+\u0026thinsp;platform\u0026thinsp;+\u0026thinsp;protocol-mean speed) bundle defined above as a single evaluation unit. Terrain and protocol speed are therefore jointly varied; full \u0026ldquo;speed-only\u0026rdquo; decoupling is out of scope. Additional file 1 provides the supplementary index mapping all Supplementary Figures and Tables to file locations. Additional files 2\u0026ndash;8 provide, respectively, domain-level permutation importance, transfer-cell source tables, condition-label mismatch sensitivity, calibration source tables, stride-time-matched speed-confound control, subject-level/biomechanical/comparator tables, and event/F18-F28 sensitivity plus endpoint-hierarchy and empirical-null robustness analyses. Partial decoupling via stride-time matching is reported in Additional file 6.\u003c/p\u003e \u003cp\u003eBox 1. Minimum in-body PE active-assistance descriptor.\u003c/p\u003e \u003cp\u003e \u003cdiv class=\"gridtable\"\u003e\u003ctable float=\"No\" id=\"Taba\" border=\"1\"\u003e \u003ccolgroup cols=\"2\"\u003e \u003cdiv align=\"left\" class=\"colspec\" colname=\"c1\" colnum=\"1\"\u003e\u003c/div\u003e \u003cdiv align=\"left\" class=\"colspec\" colname=\"c2\" colnum=\"2\"\u003e\u003c/div\u003e \u003cthead\u003e \u003ctr\u003e \u003cth align=\"left\" colname=\"c1\"\u003e \u003cp\u003eField\u003c/p\u003e \u003c/th\u003e \u003cth align=\"left\" colname=\"c2\"\u003e \u003cp\u003eIn-body PE descriptor\u003c/p\u003e \u003c/th\u003e \u003c/tr\u003e \u003c/thead\u003e \u003ctbody\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003eWearable platform\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c2\"\u003e \u003cp\u003eCable-driven soft suit with a back-mounted actuator unit and strap-fixed thigh/pelvic anchor interfaces.\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003eUE-vs-PE protocol contrast\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c2\"\u003e \u003cp\u003eSame worn hardware in both states; PE\u0026thinsp;=\u0026thinsp;actuator assistance enabled, UE\u0026thinsp;=\u0026thinsp;assistance disabled.\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003eAssistance target and direction\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c2\"\u003e \u003cp\u003eBilateral hip flexion/extension-related assistance under fixed gait-synchronous timing.\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003eAnchor path convention\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c2\"\u003e \u003cp\u003eMotor-driven cable lines were routed through strap-fixed body-surface anchor regions (force-transfer and routing/guide functions).\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003ePublic benchmark predictors\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c2\"\u003e \u003cp\u003eProcessed sEMG features and protocol metadata only.\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003eInterpretation boundary\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c2\"\u003e \u003cp\u003ePE is a categorical active-assistance protocol state; actuator-level and biomechanical detail remains outside the public benchmark scope (see Supplementary Table S0 / Additional file 1).\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003c/tbody\u003e \u003c/colgroup\u003e \u003c/table\u003e\u003c/div\u003e \u003c/p\u003e \u003cp\u003eBox 1 summarizes only the minimum in-body PE protocol descriptor used for main-text interpretation. The full release-boundary and non-claim detail is provided in Supplementary Table S0 and Additional file 1.\u003c/p\u003e\n\u003ch3\u003eEMG acquisition\u003c/h3\u003e\n\u003cp\u003eTen-channel sEMG was collected from gastrocnemius medialis (GM), tibialis anterior (TA), rectus femoris (RF), tensor fasciae latae (TFL), and biceps femoris (BF) bilaterally with the Delsys Trigno surface EMG system at 2,148 Hz, following SENIAM electrode-placement guidelines [\u003cspan citationid=\"CR38\" class=\"CitationRef\"\u003e38\u003c/span\u003e].\u003c/p\u003e\n\u003ch3\u003ePreprocessing\u003c/h3\u003e\n\u003cp\u003eEach recording was preprocessed per channel with the following steps applied in order: non-zero-sample-based outlier clipping at the 0.1\u0026ndash;99.9 percentiles (computed only when \u0026ge;\u0026thinsp;100 non-zero samples were available; zeros were excluded from the percentile statistics but retained in the signal), a 4th-order Butterworth band-pass filter (20\u0026ndash;450 Hz), a 60 Hz notch filter (Q\u0026thinsp;=\u0026thinsp;30), full-wave rectification, and Gaussian envelope smoothing (σ\u0026thinsp;=\u0026thinsp;100 ms).\u003c/p\u003e \u003cdiv id=\"Sec8\" class=\"Section2\"\u003e \u003ch2\u003eGait segmentation\u003c/h2\u003e \u003cp\u003eThroughout this manuscript we use EMG-derived HS/TO proxies rather than externally validated gait events. An EMG-derived heel-strike (HS) proxy was detected as a TA-envelope peak; an EMG-derived toe-off (TO) proxy was detected as a GM-envelope peak (scipy.signal.find_peaks; minimum distance 0.5 s). For peak-detection robustness, envelopes were normalized by median/MAD and peaks were searched stepwise with relaxed (height, prominence) = (0.5, 0.3) \u0026rarr; (0.3, 0.2) \u0026rarr; (0.1, 0.1) \u0026rarr; (0.05, 0.05) \u0026rarr; (None, 0.02) (terminating at \u0026ge;\u0026thinsp;5 peaks); the final pass used minimum distance only. A side with fewer than 2 HS peaks was excluded. HS/TO and HS\u0026rarr;next-HS cycles were constructed independently for left and right; per-side cycles were then pooled. For each cycle, TO was selected as the GM peak nearest 62% of the cycle length after HS; if absent, TO was estimated at 62%. Cycles were divided into eight fixed-percentage cycle bins based on Perry timing references [\u003cspan citationid=\"CR39\" class=\"CitationRef\"\u003e39\u003c/span\u003e]: IC (0\u0026ndash;2%), LR (2\u0026ndash;12%), MSt (12\u0026ndash;31%), TSt (31\u0026ndash;50%), PSw (50\u0026ndash;62%), ISw (62\u0026ndash;75%), MSw (75\u0026ndash;87%), and TSw (87\u0026ndash;100%). These bins are fixed-percentage cycle bins within each HS\u0026rarr;next-HS cycle and are not externally validated biomechanical gait phases. Phase boundaries used Perry\u0026rsquo;s fixed percentages and were not shifted by the detected TO. The detected TO proxy was retained only as a reference event for cycle characterization and robustness checks; the fixed-percentage bins used for classification were assigned by percentages within each HS\u0026rarr;next-HS cycle and were mapped to sample indices within that interval (no time resampling). Cycles shorter than 50 samples were excluded. Because gait-cycle boundaries and fixed-percentage phase bins were derived from EMG envelopes rather than external kinematic/force-plate events, this benchmark evaluates recognition under an EMG-derived segmentation convention. It should not be interpreted as a validation of biomechanical gait-phase detection.\u003c/p\u003e \u003cp\u003eEvent-proxy QC summary is provided in Additional file 8: sides failing minimum-peak criteria were excluded before classification, retained cycle durations were constrained by the cycle-construction filters described above, and plausible event/phase perturbation sensitivity checks did not change the primary transfer-cell ordering. Across the full collected dataset after cycle-construction filters, 29,993 cycles were retained for downstream benchmark analyses (Table\u0026nbsp;\u003cspan refid=\"Tab1\" class=\"InternalRef\"\u003e1\u003c/span\u003ea).\u003c/p\u003e \u003cp\u003e \u003cdiv class=\"gridtable\"\u003e\u003ctable float=\"Yes\" id=\"Tab1\" border=\"1\"\u003e \u003ccaption language=\"En\"\u003e \u003cdiv class=\"CaptionNumber\"\u003eTable 1\u003c/div\u003e \u003cdiv class=\"CaptionContent\"\u003e \u003cp\u003e\u003cb\u003eDataset distribution and primary cycle-level LOSO benchmark.\u003c/b\u003e a) Gait-cycle distribution by protocol condition and suit state (full collected data).\u003c/p\u003e \u003c/div\u003e \u003c/caption\u003e \u003ccolgroup cols=\"5\"\u003e \u003cdiv align=\"left\" class=\"colspec\" colname=\"c1\" colnum=\"1\"\u003e\u003c/div\u003e \u003cdiv align=\"char\" char=\".\" class=\"colspec\" colname=\"c2\" colnum=\"2\"\u003e\u003c/div\u003e \u003cdiv align=\"char\" char=\".\" class=\"colspec\" colname=\"c3\" colnum=\"3\"\u003e\u003c/div\u003e \u003cdiv align=\"char\" char=\".\" class=\"colspec\" colname=\"c4\" colnum=\"4\"\u003e\u003c/div\u003e \u003cdiv align=\"char\" char=\".\" class=\"colspec\" colname=\"c5\" colnum=\"5\"\u003e\u003c/div\u003e \u003cthead\u003e \u003ctr\u003e \u003cth align=\"left\" colname=\"c1\"\u003e \u003cp\u003eProtocol condition\u003c/p\u003e \u003c/th\u003e \u003cth align=\"left\" colname=\"c2\"\u003e \u003cp\u003eNW\u003c/p\u003e \u003c/th\u003e \u003cth align=\"left\" colname=\"c3\"\u003e \u003cp\u003eUE\u003c/p\u003e \u003c/th\u003e \u003cth align=\"left\" colname=\"c4\"\u003e \u003cp\u003ePE\u003c/p\u003e \u003c/th\u003e \u003cth align=\"left\" colname=\"c5\"\u003e \u003cp\u003eTotal\u003c/p\u003e \u003c/th\u003e \u003c/tr\u003e \u003c/thead\u003e \u003ctbody\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003eLevel\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c2\"\u003e \u003cp\u003e5583\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c3\"\u003e \u003cp\u003e5565\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c4\"\u003e \u003cp\u003e5307\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c5\"\u003e \u003cp\u003e16455\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003eSlope (18\u0026deg;)\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c2\"\u003e \u003cp\u003e2034\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c3\"\u003e \u003cp\u003e2124\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c4\"\u003e \u003cp\u003e2254\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c5\"\u003e \u003cp\u003e6412\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003eStairs\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c2\"\u003e \u003cp\u003e2290\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c3\"\u003e \u003cp\u003e2409\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c4\"\u003e \u003cp\u003e2427\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c5\"\u003e \u003cp\u003e7126\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003eTotal\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c2\"\u003e \u003cp\u003e9907\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c3\"\u003e \u003cp\u003e10098\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c4\"\u003e \u003cp\u003e9988\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c5\"\u003e \u003cp\u003e29993\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003c/tbody\u003e \u003c/colgroup\u003e \u003ctfoot\u003e \u003ctr\u003e\u003ctd colspan=\"5\"\u003eNW, not worn; UE, worn-inactive suit; PE, worn-active suit. The three protocol-defined walking conditions were performed on different platforms (only the level condition used a treadmill): level treadmill, 0.83 m/s; 18\u0026deg; slope on a 25 m walkway (50 m round trip, self-selected; trial mean 0.227 m/s); stairs on an Egojin Highclimb stepmill (step height 21 cm), 0.54 m/s. This sub-table reports the full collected distribution (29,993 cycles total) and is independent of the balanced LOSO evaluation set (\u0026le;\u0026thinsp;50 cycles per recording file).\u003c/td\u003e\u003c/tr\u003e \u003ctr\u003e\u003ctd colspan=\"5\"\u003eb) Cycle-level (primary endpoint): 3 \u0026times; 3 protocol-condition transfer matrix, where each condition bundles terrain, platform, and protocol speed (LOSO macro-F1; mean\u0026thinsp;\u0026plusmn;\u0026thinsp;95% subject-bootstrap CI; n\u0026thinsp;=\u0026thinsp;10).\u003c/td\u003e\u003c/tr\u003e \u003c/tfoot\u003e \u003c/table\u003e\u003c/div\u003e \u003c/p\u003e\u003ctable border=\"0\" cellspacing=\"0\" cellpadding=\"0\" width=\"97%\" class=\"fr-table-selection-hover\"\u003e\n \u003cthead\u003e\n \u003ctr\u003e\n \u003ctd valign=\"bottom\" style=\"width: 108px;\"\u003e\n \u003cp\u003eTrain \u0026darr; \\ Test \u0026rarr;\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"bottom\" style=\"width: 166px;\"\u003e\n \u003cp\u003eLevel (treadmill, 0.83 m/s)\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"bottom\" style=\"width: 166px;\"\u003e\n \u003cp\u003eSlope (walkway 18\u0026deg;, 0.227 m/s)\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"bottom\" style=\"width: 166px;\"\u003e\n \u003cp\u003eStairs (stepmill, 0.54 m/s)\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003c/thead\u003e\n \u003ctbody\u003e\n \u003ctr\u003e\n \u003ctd valign=\"top\" style=\"width: 108px;\"\u003e\n \u003cp\u003eLevel\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 166px;\"\u003e\n \u003cp\u003e0.43 [0.34, 0.51]\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 166px;\"\u003e\n \u003cp\u003e0.21 [0.15, 0.28]\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 166px;\"\u003e\n \u003cp\u003e0.18 [0.12, 0.25]\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd valign=\"top\" style=\"width: 108px;\"\u003e\n \u003cp\u003eSlope\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 166px;\"\u003e\n \u003cp\u003e0.24 [0.17, 0.32]\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 166px;\"\u003e\n \u003cp\u003e0.41 [0.33, 0.49]\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 166px;\"\u003e\n \u003cp\u003e0.41 [0.33, 0.49]\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd valign=\"top\" style=\"width: 108px;\"\u003e\n \u003cp\u003eStairs\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 166px;\"\u003e\n \u003cp\u003e0.23 [0.17, 0.30]\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 166px;\"\u003e\n \u003cp\u003e0.37 [0.30, 0.45]\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 166px;\"\u003e\n \u003cp\u003e0.42 [0.34, 0.50]\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003c/tbody\u003e\n\u003c/table\u003e\n\u003cp\u003eCycle-level macro-F1 = mean of available retained fixed-percentage bin prediction probabilities of one HS\u0026rarr;next-HS cycle (after NL10 filtering), then argmax. Diagonal = within-condition; off-diagonal = protocol-defined terrain/platform/speed bundle transfer. This table is the primary endpoint because the operational decision unit is one decision per gait cycle. CIs are 2,000 subject bootstraps; per-subject raw values and cell-wise confidence intervals are provided in Additional file 3. Analytical references for interpretation are class-prior-matched random \u0026asymp; 0.333 and global-majority \u0026asymp; 0.168; empirical dummy/permutation references are reported in Supplementary Table S14 and used for null interpretation.\u003c/p\u003e \u003cp\u003e \u003cdiv class=\"gridtable\"\u003e\u003ctable float=\"Yes\" id=\"Tab2\" border=\"1\"\u003e \u003ccaption language=\"En\"\u003e \u003cdiv class=\"CaptionNumber\"\u003eTable 2\u003c/div\u003e \u003cdiv class=\"CaptionContent\"\u003e \u003cp\u003eEvaluation regimes (main-text summary).\u003c/p\u003e \u003c/div\u003e \u003c/caption\u003e \u003ccolgroup cols=\"4\"\u003e \u003cdiv align=\"left\" class=\"colspec\" colname=\"c1\" colnum=\"1\"\u003e\u003c/div\u003e \u003cdiv align=\"left\" class=\"colspec\" colname=\"c2\" colnum=\"2\"\u003e\u003c/div\u003e \u003cdiv align=\"left\" class=\"colspec\" colname=\"c3\" colnum=\"3\"\u003e\u003c/div\u003e \u003cdiv align=\"left\" class=\"colspec\" colname=\"c4\" colnum=\"4\"\u003e\u003c/div\u003e \u003cthead\u003e \u003ctr\u003e \u003cth align=\"left\" colname=\"c1\"\u003e \u003cp\u003eRegime\u003c/p\u003e \u003c/th\u003e \u003cth align=\"left\" colname=\"c2\"\u003e \u003cp\u003eAllowed test-time information\u003c/p\u003e \u003c/th\u003e \u003cth align=\"left\" colname=\"c3\"\u003e \u003cp\u003eEndpoint\u003c/p\u003e \u003c/th\u003e \u003cth align=\"left\" colname=\"c4\"\u003e \u003cp\u003eRole\u003c/p\u003e \u003c/th\u003e \u003c/tr\u003e \u003c/thead\u003e \u003ctbody\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003e1. Within-condition LOSO\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c2\"\u003e \u003cp\u003enone\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c3\"\u003e \u003cp\u003ecycle-level macro-F1\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c4\"\u003e \u003cp\u003ePrimary subject-independent reference within one protocol condition\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003e2. 3\u0026times;3 protocol-condition transfer (LOSO)\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c2\"\u003e \u003cp\u003enone\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c3\"\u003e \u003cp\u003ecycle-level macro-F1 (with cycle\u0026times;phase sensitivity readout)\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c4\"\u003e \u003cp\u003ePrimary transfer-boundary reference across bundled protocol conditions\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003e3. Known-condition reference normalization\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c2\"\u003e \u003cp\u003eprotocol-condition label only (no held-out suit-state labels)\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c3\"\u003e \u003cp\u003ecycle\u0026times;phase delta macro-F1 / ratio change versus Regime 2\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c4\"\u003e \u003cp\u003eSecondary known-condition calibration sensitivity\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003e4. Offline pooled subject-window normalization\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c2\"\u003e \u003cp\u003efirst 4 unlabeled cycles per held-out subject and condition\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c3\"\u003e \u003cp\u003ecycle\u0026times;phase macro-F1 on remaining cycles\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c4\"\u003e \u003cp\u003eSecondary unlabeled adaptation sensitivity (offline reference)\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003e5. Small-label post-hoc calibration (LOTO under LOSO)\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c2\"\u003e \u003cp\u003esmall labeled cycle budget (phase-label-equivalent reporting)\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c3\"\u003e \u003cp\u003ecycle\u0026times;phase macro-F1\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c4\"\u003e \u003cp\u003eSecondary label-budget recovery characterization (different regime)\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003e6. Latency benchmark\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c2\"\u003e \u003cp\u003en/a\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c3\"\u003e \u003cp\u003einference-only and end-to-end latency, reported separately\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c4\"\u003e \u003cp\u003eComputational-cost readout (not deployment-readiness claim)\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003e7. Asymmetric critical-error readout\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c2\"\u003e \u003cp\u003en/a\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c3\"\u003e \u003cp\u003eP(pred\u0026thinsp;=\u0026thinsp;PE|true\u0026thinsp;=\u0026thinsp;NW) and P(pred\u0026thinsp;=\u0026thinsp;NW|true\u0026thinsp;=\u0026thinsp;PE)\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c4\"\u003e \u003cp\u003eRisk visibility beyond macro-F1 within the same regime\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003c/tbody\u003e \u003c/colgroup\u003e \u003ctfoot\u003e \u003ctr\u003e\u003ctd colspan=\"4\"\u003eFull regime definitions and supplementary index mapping are provided in Additional file 1.\u003c/td\u003e\u003c/tr\u003e \u003c/tfoot\u003e \u003c/table\u003e\u003c/div\u003e \u003c/p\u003e \u003c/div\u003e\n\u003ch3\u003eFeature extraction\u003c/h3\u003e\n\u003cp\u003eF18 (18-D) consisted of time-domain (6), frequency-domain (5), and wavelet-domain (7) features. Event detection was performed on the envelope; F18 features were computed on the band-passed signal (bp) and its envelope (env): RMS / MAV / variance from env; waveform length (WL), zero-crossing rate (ZCR), slope sign change (SSC) from bp (sign-change-based, no ε threshold). Frequency features used Welch\u0026rsquo;s PSD on bp (scipy.signal.welch; nperseg\u0026thinsp;=\u0026thinsp;min(L, 1024); other parameters at SciPy defaults), restricted to 20\u0026ndash;450 Hz; relative band power was defined as: integral of P(f) over the target band divided by integral of P(f) over 20\u0026ndash;450 Hz. To avoid target-label dependence, wavelet features used a condition-agnostic fixed 3-wavelet ensemble (level\u0026thinsp;=\u0026thinsp;5): {db5, sym6, coif3} with fixed weights {0.5, 0.3, 0.2}. NL10 (10-D) included sample entropy (m\u0026thinsp;=\u0026thinsp;2, r\u0026thinsp;=\u0026thinsp;0.2; z-scored signal), approximate entropy (m\u0026thinsp;=\u0026thinsp;2, r\u0026thinsp;=\u0026thinsp;0.2), permutation entropy (order\u0026thinsp;=\u0026thinsp;3, delay\u0026thinsp;=\u0026thinsp;1, normalized), Higuchi FD (kmax\u0026thinsp;=\u0026thinsp;min(10, N/4)), Katz FD, Hurst (max_lag\u0026thinsp;=\u0026thinsp;min(50, N/4)), recurrence-quantification metrics (m\u0026thinsp;=\u0026thinsp;2, τ\u0026thinsp;=\u0026thinsp;1, threshold\u0026thinsp;=\u0026thinsp;0.1\u0026times; max distance; DET lmin\u0026thinsp;=\u0026thinsp;2; LAM vmin\u0026thinsp;=\u0026thinsp;2), and the Rosenstein largest Lyapunov exponent (m\u0026thinsp;=\u0026thinsp;2, τ\u0026thinsp;=\u0026thinsp;1). NL10 was computed only on segments\u0026thinsp;\u0026ge;\u0026thinsp;50 samples (\u0026asymp;\u0026thinsp;23 ms at 2,148 Hz); shorter segments were NaN, and any cycle\u0026times;phase sample with NaN was excluded from F28 analyses (Supplementary Table \u003cspan refid=\"MOESM3\" class=\"InternalRef\"\u003eS3\u003c/span\u003e). Feature-block definitions are summarized in Additional file 1 Table A3.\u003c/p\u003e \u003cp\u003eWavelet 7-D terms were computed from the same level-5 decomposition for each basis w in {db5, sym6, coif3}: WE_w\u0026thinsp;=\u0026thinsp;sum_k c_w,k^2 (detail-coefficient energy), WEnt_w = -sum_k p_w,k log p_w,k with p_w,k\u0026thinsp;=\u0026thinsp;c_w,k^2 / sum_j c_w,j^2, and cross-wavelet ratio = (WE_db5\u0026thinsp;+\u0026thinsp;WE_sym6) / (WE_coif3\u0026thinsp;+\u0026thinsp;1e-12). These descriptors are reported as computational features for classification and not as direct mechanistic biomarkers.\u003c/p\u003e\n\u003ch3\u003ePredefined literature-based per-muscle scaling convention\u003c/h3\u003e\n\u003cp\u003ePer-muscle scaling weights for muscle m were defined as follows:\u003c/p\u003e \u003cp\u003ew_raw,m\u0026thinsp;=\u0026thinsp;w_PCSA,m x w_FT,m x w_pen,m;\u003c/p\u003e \u003cp\u003ew_PCSA,m\u0026thinsp;=\u0026thinsp;1 / PCSA_m;\u003c/p\u003e \u003cp\u003ew_FT,m\u0026thinsp;=\u0026thinsp;1\u0026thinsp;+\u0026thinsp;0.6 x (FT_m\u0026thinsp;\u0026minus;\u0026thinsp;0.5);\u003c/p\u003e \u003cp\u003ew_pen,m\u0026thinsp;=\u0026thinsp;1 - theta_m / 90 deg;\u003c/p\u003e \u003cp\u003ew_m\u0026thinsp;=\u0026thinsp;w_raw,m / mean(w_raw across muscles);\u003c/p\u003e \u003cp\u003ePhi_B\u0026thinsp;=\u0026thinsp;diag(w).\u003c/p\u003e \u003cp\u003ewhere PCSA is physiological cross-sectional area, FT is fast-fibre fraction, and theta is pennation angle. The normalization sets the across-muscle mean of w to 1.0. Phi_B is a diagonal per-muscle scalar applied to all extracted features and is not personalized per subject. Muscle parameters (PCSA, FT, theta) were taken from the anatomical literature [\u003cspan citationid=\"CR40\" class=\"CitationRef\"\u003e40\u003c/span\u003e, \u003cspan citationid=\"CR41\" class=\"CitationRef\"\u003e41\u003c/span\u003e] as fixed per-muscle source-value triplets (left\u0026thinsp;=\u0026thinsp;right), with the resulting Phi_B weights reported in Supplementary Table \u003cspan refid=\"MOESM7\" class=\"InternalRef\"\u003eS7\u003c/span\u003e. Phi_B was retained only as a frozen, label-independent benchmark preprocessing convention, not because it improved performance or because it represents a validated subject-specific biomechanical model. Because Phi_B did not show a uniform performance gain, all primary transfer-boundary conclusions were checked against Phi_B-off comparators, and no claim depends on Phi_B improving recognition performance.\u003c/p\u003e \u003cdiv id=\"Sec11\" class=\"Section2\"\u003e \u003ch2\u003eNo-leakage design summary\u003c/h2\u003e \u003cp\u003e \u003cul\u003e \u003cli\u003e \u003cp\u003e\u0026bull; One analysis sample is one cycle\u0026times;phase 28-D vector (10-channel mean after per-muscle scaling).\u003c/p\u003e \u003c/li\u003e \u003cli\u003e \u003cp\u003e\u0026bull; Primary endpoint is cycle-level macro-F1 (mean of available retained phase probabilities per HS\u0026rarr;next-HS cycle).\u003c/p\u003e \u003c/li\u003e \u003cli\u003e \u003cp\u003e\u0026bull; LOSO split holds out one subject entirely; no held-out-subject labels are used in model fitting.\u003c/p\u003e \u003c/li\u003e \u003cli\u003e \u003cp\u003e\u0026bull; In this manuscript, leakage-controlled means that held-out-subject suit-state labels, held-out-subject class information, and held-out-subject feature statistics were not used for model fitting, normalization, calibration, feature scaling, or prediction. Ground-truth labels were used only after prediction for pre-specified evaluation-set definition and metric computation.\u003c/p\u003e \u003c/li\u003e \u003cli\u003e \u003cp\u003e\u0026bull; Signal-level preprocessing and EMG-event proxy construction are offline within-recording feature-generation steps, not causal online deployment procedures.\u003c/p\u003e \u003c/li\u003e \u003cli\u003e \u003cp\u003e\u0026bull; Balanced LOSO evaluation uses\u0026thinsp;\u0026le;\u0026thinsp;50 cycles per recording file (recording file\u0026thinsp;=\u0026thinsp;one subject-specific trial/run under one protocol condition and one suit state) to avoid dominance by long files.\u003c/p\u003e \u003c/li\u003e \u003cli\u003e \u003cp\u003e\u0026bull; Known-condition reference normalization uses training-subject-only (mu, sigma) by protocol-defined condition and requires the correct test protocol-condition label.\u003c/p\u003e \u003c/li\u003e \u003cli\u003e \u003cp\u003e\u0026bull; The pooled subject-window analysis was retained as an offline reference normalization sensitivity analysis because the calibration window depended on recording-file structure and was excluded from evaluation.\u003c/p\u003e \u003c/li\u003e \u003cli\u003e \u003cp\u003e\u0026bull; Class-stratified post-hoc calibration characterization samples labels at whole-cycle level and excludes the sampled cycles in full.\u003c/p\u003e \u003c/li\u003e \u003c/ul\u003e \u003c/p\u003e \u003c/div\u003e \u003cdiv id=\"Sec12\" class=\"Section2\"\u003e \u003ch2\u003eClassification\u003c/h2\u003e \u003cp\u003eThe benchmark hierarchy was defined before reporting: cycle-level macro-F1 under LOSO was the primary endpoint; all cycle\u0026times;phase, calibration, latency, coordination, and ablation analyses were secondary or supporting readouts. A Random Forest classifier (RF; scikit-learn 1.7.1 [\u003cspan citationid=\"CR42\" class=\"CitationRef\"\u003e42\u003c/span\u003e]; n_estimators\u0026thinsp;=\u0026thinsp;200, random_state\u0026thinsp;=\u0026thinsp;42) [\u003cspan citationid=\"CR43\" class=\"CitationRef\"\u003e43\u003c/span\u003e] was used under LOSO with a three-class target (NW, UE, PE). Each cycle\u0026times;phase sample was a single 28-D vector \u0026mdash; the 10-channel arithmetic mean of the per-channel F28 features (F28\u0026thinsp;=\u0026thinsp;F18\u0026thinsp;+\u0026thinsp;NL10) after the frozen per-muscle preprocessing convention (Φ_B), treated as a predefined label-independent convention rather than a learned performance-enhancement step. The cycle-level prediction is the mean of the available retained phase probabilities within one HS\u0026rarr;next-HS cycle and is the primary endpoint; cycle\u0026times;phase numbers are reported only as a finer-grained sensitivity readout.\u003c/p\u003e \u003cp\u003eTraining unit vs. endpoint unit. The primary model is trained on phase-level samples (8 phase bins per cycle), but the operational decision unit and primary endpoint are at the cycle level (one decision per HS\u0026rarr;next-HS cycle). We retain phase-level training because the predefined benchmark implementation aggregates phase-wise class probabilities to one cycle-level decision while preserving phase-resolved sensitivity readouts (Fig.\u0026nbsp;\u003cspan refid=\"Fig2\" class=\"InternalRef\"\u003e2\u003c/span\u003eb). Because phase samples within one cycle are correlated, primary uncertainty is computed at held-out subject level (subject bootstrap). A cycle-weighted training sensitivity (sample weights of 1 / retained-phase-count per cycle) is reported in Additional file 8 (Supplementary Table S12). As an additional robustness analysis, direct cycle-level comparators were evaluated on the same split manifest by aggregating retained phase features within each cycle using mean pooling (cycle-level 28-D vectors), followed by cycle-level RF and multinomial logistic regression under LOSO (Supplementary Table S13, Additional file 8). This design does not claim phase-level training superiority; it defines the benchmark hierarchy and its cycle-level comparators explicitly. Accordingly, phase-level training is treated as an implementation choice of the benchmark baseline, whereas all primary interpretation is tied to the cycle-level endpoint and to robustness checks that remove or down-weight within-cycle phase multiplicity.\u003c/p\u003e \u003cp\u003e \u003c/p\u003e \u003cp\u003e \u003c/p\u003e \u003cp\u003eBalanced LOSO sampling algorithm (seed\u0026thinsp;=\u0026thinsp;42). 1. Build a file list where one recording file is one (subject, protocol condition, suit state, trial/run file) unit. 2. For each file, select up to 50 HS\u0026rarr;HS cycles once using seed 42 and write selected cycle IDs to splits/loso_manifest_v1.csv. 3. Reuse the same manifest across all LOSO folds (no per-fold re-sampling). 4. Train/evaluate only on manifest-selected cycles (capped design for both training and evaluation in this benchmark definition). 5. After NL10 filtering, retain a cycle for cycle-level scoring if at least one phase probability remains; compute cycle-level probabilities as the arithmetic mean of retained phase probabilities.\u003c/p\u003e \u003cp\u003eFor F28 analyses, cycle-level aggregation was computed from retained phase-level F28 entries after NL10 filtering (Supplementary Table \u003cspan refid=\"MOESM3\" class=\"InternalRef\"\u003eS3\u003c/span\u003e); cycle-level denominators therefore correspond to cycles with available phase probabilities under this filter.\u003c/p\u003e \u003c/div\u003e \u003cdiv id=\"Sec13\" class=\"Section2\"\u003e \u003ch2\u003eSecondary calibration analyses\u003c/h2\u003e \u003cp\u003eKnown-condition reference normalization. Condition-specific z-score normalization was computed from training subjects only (train_only_by_condition). For each feature dimension x, x\u0026rsquo; = (x\u0026thinsp;\u0026minus;\u0026thinsp;\u0026micro;_{train, condition}) / σ_{train, condition}; \u0026micro; and σ were estimated from training subjects only and selected according to the protocol-defined condition label of each sample. No held-out-subject statistics or labels were used. For a train-condition \u0026rarr; test-protocol-condition transfer cell, classifier fitting used only train-condition samples from training subjects, whereas test-condition normalization statistics were estimated from training-subject recordings in that same test condition. This regime is therefore a known-condition reference setting, not a condition-agnostic unsupervised adaptation setting.\u003c/p\u003e \u003cp\u003eOffline pooled subject-window normalization analysis. In the held-out subject, the first 4 cycles per recording file per test condition were separated as an unlabelled calibration window and excluded from evaluation. Calibration-window cycles were pooled across recording files within each held-out subject \u0026times; test condition cell to estimate one normalization vector. No suit-state-specific or file-specific normalization statistics were estimated or applied. Because the window is defined by recording-file structure, this analysis is retained as an offline reference normalization sensitivity analysis rather than an autonomous deployment policy.\u003c/p\u003e \u003cp\u003eClass-stratified post-hoc calibration characterization. In leave-one-protocol-condition-out under LOSO, two protocol-defined conditions were used for training and the held-out condition for testing. Known-condition reference normalization was applied first, and a multinomial logistic regression was then fitted post-hoc to log-RF-probability outputs in the held-out subject \u0026times; test condition. Labels were sampled and excluded at the whole-cycle level. Budgets are reported as phase-label-equivalent units (10/30/90), and the procedure was repeated 30 times with random cycle sampling (seed\u0026thinsp;=\u0026thinsp;42). This LOTO-under-LOSO regime is a secondary characterization only and is excluded from primary-endpoint effectiveness claims.\u003c/p\u003e \u003c/div\u003e \u003cdiv id=\"Sec14\" class=\"Section2\"\u003e \u003ch2\u003eUncertainty estimation and supporting statistical analyses\u003c/h2\u003e \u003cp\u003e95% confidence intervals were obtained by 2,000 subject bootstraps (resampling subjects with replacement; 2.5\u0026ndash;97.5 percentile range). Macro-F1 was computed per held-out subject and aggregated as the unweighted (subject) mean; CIs were computed on these subject-level scores. For paired regime-comparison deltas, Δ macro-F1 was computed first within held-out subject, and 95% CIs were obtained by bootstrapping the paired subject-level differences with the same 2,000 subject-resampling procedure. Domain-level group permutation importance was computed under in-condition LOSO (per protocol condition and per held-out subject; baseline macro-F1 vs the macro-F1 after permuting each domain\u0026rsquo;s feature columns 20 times in the held-out subject\u0026rsquo;s test set; 95% CIs by 2,000 bootstraps over per-subject mean drops). Coordination metrics used per-subject, per (protocol condition \u0026times; suit state) cell 10\u0026times;10 Pearson correlation matrices of per-muscle 28-D feature trajectories, with mean |r|, mean r, and the network-density AUC over the threshold range (trapezoidal integration; step\u0026thinsp;=\u0026thinsp;0.01). End-to-end latency was measured on Apple M1 across 5 cycles per protocol condition (N\u0026thinsp;=\u0026thinsp;15 cycles total): event detection was timed once per file and divided by cycle count; per-cycle processing comprised preprocessing, phase indexing, feature extraction (F18 or F28), and inference. Inference-only latency used pre-computed features (5 warm-up +\u0026thinsp;1,000 measured calls). Environment identifiers were recorded (macOS 26.3 arm64; Python 3.11.5; numpy 1.26.4; scipy 1.15.2; scikit-learn 1.7.1; PyWavelets 1.8.0; commit 4e38a5b409e338c1e890bb64f86461fa4cd46926). RF inference used scikit-learn parallelism (n_jobs\u0026thinsp;=\u0026thinsp;\u0026minus;\u0026thinsp;1). The \u0026ldquo;~9 Hz throughput\u0026rdquo; mentioned in the body is the inverse of the F18\u0026thinsp;+\u0026thinsp;RF mean end-to-end latency (\u0026asymp;\u0026thinsp;1/0.115 s\u0026thinsp;\u0026asymp;\u0026thinsp;8.7 Hz). Event-detection sensitivity varied σ \u0026isin; {50, 100, 150} ms and min_interval \u0026isin; {0.4, 0.5, 0.6} s; classification sensitivity used artificial phase-boundary shifts of \u0026plusmn;\u0026thinsp;10/\u0026plusmn;20/\u0026plusmn;40 ms with F18 recomputed. Paired two-tailed t-tests with Benjamini\u0026ndash;Hochberg correction (q\u0026thinsp;\u0026lt;\u0026thinsp;0.05) were used for ablation analyses on subject-level (n\u0026thinsp;=\u0026thinsp;10) within-subject paired metrics. Empirical null analyses were computed on the same balanced LOSO split manifest at the cycle-level endpoint. Dummy baselines used scikit-learn DummyClassifier outputs for most-frequent and stratified strategies, and a uniform-probability dummy evaluated through the same probability-aggregation and deterministic-argmax scoring pipeline, followed by the same cycle-level macro-F1 scoring and subject-level aggregation as the primary endpoint. Permutation-null references were generated at cycle level by permuting true cycle labels within each held-out-subject \u0026times; train-condition \u0026times; test-condition cell while holding model predictions fixed, with 500 repetitions per cell; null summaries are reported in Supplementary Table S14.\u003c/p\u003e \u003cp\u003eFor the abstention analysis (Fig.\u0026nbsp;\u003cspan refid=\"Fig3\" class=\"InternalRef\"\u003e4\u003c/span\u003ea), the confidence score was the cycle-level maximum class probability. Predictions with confidence below a threshold τ were abstained; coverage was defined as the fraction of non-abstained cycles. The error curve reports the PE\u0026larr;NW pre-specified asymmetric error rate among non-abstained off-diagonal transfer-cell predictions (PE\u0026larr;NW denotes the error in which a true NW cycle is classified as PE). The threshold τ was swept across the observed confidence range and the resulting curves were aggregated across held-out subjects.\u003c/p\u003e \u003cp\u003e \u003c/p\u003e \u003c/div\u003e \u003cdiv id=\"Sec15\" class=\"Section2\"\u003e \u003ch2\u003eSupporting analyses\u003c/h2\u003e \u003cp\u003eNMF activation-pattern visualization, coordination-network metrics, feature-module ablation, abstention analysis, and event-detection sensitivity analyses were retained as supporting or robustness readouts. These analyses document descriptive signal structure and processing sensitivity; they do not define the primary endpoint and are not interpreted as mechanistic evidence of altered muscle synergies or closed-loop control behavior.\u003c/p\u003e \u003cp\u003eThe benchmark is intended to test whether a future method can exceed the fixed null-referenced LOSO transfer boundary under the same manifest and scoring endpoint, not to establish that the current RF baseline is an adequate recognizer.\u003c/p\u003e \u003c/div\u003e \u003cdiv id=\"Sec16\" class=\"Section2\"\u003e \u003ch2\u003eUse of AI tools\u003c/h2\u003e \u003cp\u003eChatGPT (OpenAI) was used only for non-substantive language polishing and literature-search query formulation. All references, claims, analyses, numerical results, figures, and tables were selected, generated, verified, and approved by the authors. No LLM was listed as an author.\u003c/p\u003e \u003c/div\u003e"},{"header":"Results","content":"\u003cdiv id=\"Sec18\" class=\"Section2\"\u003e \u003ch2\u003eDataset and primary LOSO benchmark\u003c/h2\u003e \u003cp\u003eFrom 10 healthy adult men, 29,993 EMG-derived gait cycles were retained after cycle-construction filters across three protocol-defined walking-condition bundles and three externally imposed suit states. The primary endpoint was cycle-level macro-F1 under LOSO, with one held-out subject excluded entirely from model fitting and one prediction produced per HS\u0026rarr;next-HS cycle. Phase-level samples were used for model fitting, but the pre-specified operational endpoint was the cycle-level decision obtained by averaging retained phase-wise class probabilities within each cycle. Within the balanced LOSO manifest, the evaluation set was capped at \u0026le;\u0026thinsp;50 cycles per recording file across 90 recording files (3 protocol conditions \u0026times; 3 suit states \u0026times; 10 subjects), producing up to 4,500 cycle-level samples (\u0026asymp;\u0026thinsp;1,500 per protocol condition) and 12,000 cycle\u0026times;phase samples per protocol condition before NL10 filtering; after NL10 filtering, all 4,500 manifest-selected cycles contributed at least one retained phase probability and were used for cycle-level scoring; cell-wise retained counts are provided in Supplementary Table \u003cspan refid=\"MOESM3\" class=\"InternalRef\"\u003eS3\u003c/span\u003e.\u003c/p\u003e \u003cp\u003eIn the primary cycle-level endpoint, within-condition macro-F1 was 0.43 [0.34, 0.51] for level, 0.41 [0.33, 0.49] for slope, and 0.42 [0.34, 0.50] for stairs. Across protocol-condition transfers, macro-F1 ranged from 0.18 to 0.41. Thus, the main empirical finding is not high absolute recognizer performance but a clear subject-independent transfer boundary under protocol-condition changes.\u003c/p\u003e \u003c/div\u003e \u003cdiv id=\"Sec19\" class=\"Section2\"\u003e \u003ch2\u003eProtocol-condition transfer gaps\u003c/h2\u003e \u003cp\u003eThe transfer matrix should be interpreted as a protocol-condition transfer matrix, not as a pure terrain-only generalization test. The weakest transfers were level treadmill \u0026rarr; slope walkway and level treadmill \u0026rarr; stair stepmill. Secondary cycle\u0026times;phase readouts showed the same diagonal-versus-off-diagonal ordering and the same weak transfer cells, but they remain sensitivity analyses because the eight phase samples within a gait cycle are correlated. Because the primary model used phase-level training while the endpoint was cycle-level, direct cycle-level RF and multinomial logistic-regression comparators, trained and evaluated on one aggregated vector per cycle, were treated as hierarchy checks; RF reached 0.36\u0026ndash;0.44 within condition and 0.16\u0026ndash;0.44 across transfers, and multinomial logistic regression reached 0.37\u0026ndash;0.45 within condition and 0.17\u0026ndash;0.38 across transfers, while preserving the same diagonal-versus-off-diagonal ordering. Numerical comparator values are reported in Table\u0026nbsp;\u003cspan refid=\"Tab3\" class=\"InternalRef\"\u003e3\u003c/span\u003e and Supplementary Table S13. See Fig.\u0026nbsp;\u003cspan refid=\"Fig2\" class=\"InternalRef\"\u003e2\u003c/span\u003e for secondary cycle\u0026times;phase transfer-cell ordering. Findings should be interpreted as protocol-condition transfer across bundled terrain/platform/speed conditions, not as terrain-only generalization.\u003c/p\u003e \u003cp\u003e \u003cdiv class=\"gridtable\"\u003e\u003ctable float=\"Yes\" id=\"Tab3\" border=\"1\"\u003e \u003ccaption language=\"En\"\u003e \u003cdiv class=\"CaptionNumber\"\u003eTable 3\u003c/div\u003e \u003cdiv class=\"CaptionContent\"\u003e \u003cp\u003e\u003cb\u003eHeadline benchmark numbers.\u003c/b\u003e This condensed table keeps only the four headline messages for quick interpretation; detailed null-reference values and direct cycle-level comparator numerics are reported in Supplementary Tables S13\u0026ndash;S14 and Additional file 8 under the same LOSO split manifest.\u003c/p\u003e \u003c/div\u003e \u003c/caption\u003e \u003ccolgroup cols=\"3\"\u003e \u003cdiv align=\"left\" class=\"colspec\" colname=\"c1\" colnum=\"1\"\u003e\u003c/div\u003e \u003cdiv align=\"left\" class=\"colspec\" colname=\"c2\" colnum=\"2\"\u003e\u003c/div\u003e \u003cdiv align=\"left\" class=\"colspec\" colname=\"c3\" colnum=\"3\"\u003e\u003c/div\u003e \u003cthead\u003e \u003ctr\u003e \u003cth align=\"left\" colname=\"c1\"\u003e \u003cp\u003eBlock\u003c/p\u003e \u003c/th\u003e \u003cth align=\"left\" colname=\"c2\"\u003e \u003cp\u003ePurpose\u003c/p\u003e \u003c/th\u003e \u003cth align=\"left\" colname=\"c3\"\u003e \u003cp\u003eKey finding\u003c/p\u003e \u003c/th\u003e \u003c/tr\u003e \u003c/thead\u003e \u003ctbody\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003ePrimary F28 cycle-level LOSO\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c2\"\u003e \u003cp\u003emain benchmark\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c3\"\u003e \u003cp\u003ewithin 0.41\u0026ndash;0.43, transfer 0.18\u0026ndash;0.41\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003eDirect cycle-level comparator\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c2\"\u003e \u003cp\u003eendpoint hierarchy check\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c3\"\u003e \u003cp\u003eRF 0.36\u0026ndash;0.44 (within), 0.16\u0026ndash;0.44 (transfer); LR 0.37\u0026ndash;0.45 (within), 0.17\u0026ndash;0.38 (transfer); same ordering\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003eCalibration sensitivity\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c2\"\u003e \u003cp\u003esecondary only\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c3\"\u003e \u003cp\u003eweak cells improve only with known-condition reference\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003eLatency\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c2\"\u003e \u003cp\u003ecomputational cost only\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c3\"\u003e \u003cp\u003eF28 offline, F18 lower-latency comparator\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003c/tbody\u003e \u003c/colgroup\u003e \u003c/table\u003e\u003c/div\u003e \u003c/p\u003e \u003cp\u003eAnalytical references for 3-class macro-F1 were approximately 0.333 for class-prior-matched random prediction and 0.168 for global-majority prediction. Under these references, the within-condition cells were modestly above null ranges, whereas several off-diagonal transfer cells were near null.\u003c/p\u003e \u003c/div\u003e \u003cdiv id=\"Sec20\" class=\"Section2\"\u003e \u003ch2\u003eSecondary calibration readouts under separate regimes\u003c/h2\u003e \u003cp\u003eCalibration-related results are reported as regime-specific secondary readouts rather than as deployment adaptation results. In secondary cycle\u0026times;phase readouts, known-condition reference normalization increased macro-F1 in the two weakest transfer cells: level \u0026rarr; slope by +\u0026thinsp;0.095 [0.04, 0.15] and level \u0026rarr; stairs by +\u0026thinsp;0.098 [0.05, 0.15]. Because this regime requires the correct test protocol-condition label, these deltas are interpreted as oracle-style reference-normalization sensitivity rather than condition-agnostic adaptation or deployable online normalization. These values correspond to Table\u0026nbsp;\u003cspan refid=\"Tab3\" class=\"InternalRef\"\u003e3\u003c/span\u003e (Block C) and Fig.\u0026nbsp;\u003cspan refid=\"Fig4\" class=\"InternalRef\"\u003e3\u003c/span\u003e.\u003c/p\u003e \u003cp\u003e \u003c/p\u003e \u003cp\u003e The offline pooled subject-window normalization analysis provided an additional sensitivity readout using the first four unlabeled held-out-subject cycles per file/condition as an excluded calibration window. Because this window depended on recording-file structure, it is retained only as an offline reference analysis. The small-label LOTO-under-LOSO analysis characterized post-hoc calibration behavior under cycle-level label sampling and is not treated as an autonomous label-acquisition or deployment protocol. Class-wise confusion summaries in Additional file 3 indicate that the weakest transfer cells were characterized by concentrated off-diagonal confusion rather than uniform three-class degradation; therefore, the three-class results should be interpreted as protocol-state discriminability under the benchmark labels, not as isolated active-assistance detection.\u003c/p\u003e \u003c/div\u003e \u003cdiv id=\"Sec21\" class=\"Section2\"\u003e \u003ch2\u003eComputational latency comparator and supporting analyses\u003c/h2\u003e \u003cp\u003eIn the offline replay latency benchmark on Apple M1, the F18\u0026thinsp;+\u0026thinsp;RF path required 115 ms/cycle end-to-end, whereas the full F28 path required 2697 ms/cycle and remained an offline processed-feature reference. RF inference on pre-computed features required 1.36 ms on average. These numbers are computational readouts only and do not represent within-cycle closed-loop control latency. The latency analysis separates computational cost from recognition performance: the full processed-feature path is retained as the primary offline benchmark reference for discriminability profiling, whereas the lower-latency F18 path is reported as a computational comparator under the same split manifest; the two paths are not combined into a single recognizer or deployment claim. All latency measurements were obtained from an offline-replay, post-cycle computational benchmark on Apple M1; the cycle-level decision is available only after HS\u0026rarr;next-HS cycle completion. Latency constraints and abstention summaries are linked in Table\u0026nbsp;\u003cspan refid=\"Tab3\" class=\"InternalRef\"\u003e3\u003c/span\u003e (Block D) and Fig.\u0026nbsp;\u003cspan refid=\"Fig3\" class=\"InternalRef\"\u003e4\u003c/span\u003e. Event-detection and phase-boundary sensitivity analyses did not define the primary endpoint but were used to check whether the main transfer ordering was dominated by plausible HS/TO proxy perturbations; full source values are provided in Additional file 8.\u003c/p\u003e \u003cp\u003eSupporting NMF, coordination-network, feature-module ablation, abstention-policy, and event-sensitivity analyses are provided as descriptive or robustness readouts. They were not used to define the primary endpoint and are not interpreted as mechanistic evidence of altered muscle synergies, safety validation, or closed-loop control behavior.\u003c/p\u003e \u003c/div\u003e"},{"header":"Discussion","content":"\u003cp\u003eThis study contributes a reusable null-referenced transfer-boundary benchmark for EMG-only recognition modules in wearable gait-assistance research. Its value lies not in high absolute recognition accuracy, but in defining where subject-independent transfer fails under fixed, processed-feature-level reproducible LOSO scoring. The modest macro-F1 values therefore provide a reproducible reliability boundary that future recognition methods, sensor-fusion pipelines, or calibration strategies should exceed under the same split manifest and scoring procedure. The contribution is a fixed benchmark procedure\u0026mdash;split manifest, transfer grid, null references, scoring endpoint, and reproducibility package\u0026mdash;rather than a new classifier or deployment controller.\u003c/p\u003e \u003cp\u003eCalibration analyses are regime-specific secondary readouts. Known-condition reference normalization increased macro-F1 in selected weak transfer cells but required the correct test protocol-condition label. The pooled subject-window analysis remained an offline sensitivity readout because its calibration window depended on recording-file structure and was excluded from evaluation. The small-label LOTO-under-LOSO analysis characterized post-hoc behavior under cycle-level label sampling and is not a deployment policy. Because the ΦB ON-versus-OFF comparator did not show a uniform macro-F1 gain, ΦB should be interpreted only as a predefined label-independent pre-fusion scaling convention. The reported benchmark should therefore not be read as evidence that ΦB improves subject-independent recognition performance or as evidence for a mechanistic muscle-force model.\u003c/p\u003e \u003cp\u003eThe latency analysis should be read as an offline computational constraint analysis. F28 defines the higher-cost processed-feature reference used for the primary discriminability benchmark, whereas F18 defines a lower-latency computational comparator. Because the cycle-level decision is available only after HS\u0026rarr;next-HS cycle completion, and because closed-loop control was not implemented, neither latency pathway demonstrates within-cycle control readiness. F18 performance analyses are retained as supplementary comparator readouts (Supplementary Table \u003cspan refid=\"MOESM10\" class=\"InternalRef\"\u003eS10\u003c/span\u003e), while primary endpoint interpretation remains anchored to F28-based cycle-level LOSO evaluation.\u003c/p\u003e \u003cp\u003eSupporting NMF, coordination-network, feature-module ablation, abstention-policy, and event-sensitivity analyses document descriptive signal structure and robustness checks. They are not presented as mechanistic evidence of altered muscle synergies, safety validation, or controller-readiness evidence. This separation is essential because the primary benchmark is a processed-feature recognition-stage evaluation.\u003c/p\u003e \u003cp\u003eLimitations. The sample included only 10 healthy adult men and one cable-driven suit platform. Terrain, platform, and locomotor speed were jointly varied, so the transfer matrix is not a pure terrain-only test. EMG-derived event proxies were not validated against external sensors. The released benchmark is processed-feature based; raw EMG redistribution remains restricted under the stated DUA procedure. The protocol metadata did not explicitly log randomization/counterbalancing order, scheduled rest intervals, or electrode reattachment events, leaving residual order, fatigue, session, or electrode-state effects as interpretation boundaries. Finally, closed-loop control was not implemented, so no deployment-level control claim follows from these results. In addition, the anchor and cable-path descriptors define the intended assistance configuration only; body-surface anchor motion, strap-interface pressure, cable-line migration, and subject-specific delivered joint moments were not measured in this benchmark. Because phase-level samples from the same cycle are correlated, primary inference in the present report is based on subject-level uncertainty, and a cycle-weighted RF sensitivity was additionally reported as a robustness check (Additional file 8). Direct cycle-level aggregation comparators and empirical dummy/permutation null baselines were added in this revision as robustness analyses on the same LOSO split (Supplementary Tables S13-S14, Additional file 8). These analyses reduce, but do not eliminate, interpretation uncertainty around phase-level training versus cycle-level endpoint hierarchy. Because all participants were healthy adult men, sex- or gender-related variability could not be evaluated, and these findings should not be generalized to women, older adults, or clinical populations without external validation. Importantly, the released package supports reproducibility of the processed-feature recognition benchmark, but does not fully reproduce actuator-level assistance waveforms; therefore PE-related findings should be interpreted as protocol-state discriminability under a restricted assistance descriptor rather than as evidence of a specific torque-profile effect.\u003c/p\u003e"},{"header":"Conclusions","content":"\u003cp\u003eThis proof-of-concept study provides a leakage-controlled processed-feature benchmark for sEMG-based discriminability of externally imposed NW/UE/PE protocol states in a cable-driven walking-assistance suit. Under LOSO evaluation, cycle-level macro-F1 was modest within condition and degraded under protocol-condition transfer, demonstrating substantial subject-independent transfer gaps. The benchmark therefore defines a reproducible transfer-boundary and method-comparison reference rather than a deployment-ready recognition or control system. External validation across larger and more diverse populations, alternative suit platforms, decoupled speed protocols, externally validated gait events, and closed-loop experiments is required before deployment-level claims can be made.\u003c/p\u003e"},{"header":"Abbreviations","content":"\u003cdiv class=\"DefinitionList\"\u003e \u003cdiv class=\"DefinitionListEntry\"\u003e \u003cdiv class=\"Term\"\u003eApEn\u003c/div\u003e \u003cdiv class=\"Description\"\u003e \u003cp\u003eapproximate entropy\u003c/p\u003e \u003c/div\u003e \u003c/div\u003e \u003cdiv class=\"DefinitionListEntry\"\u003e \u003cdiv class=\"Term\"\u003eBF\u003c/div\u003e \u003cdiv class=\"Description\"\u003e \u003cp\u003ebiceps femoris\u003c/p\u003e \u003c/div\u003e \u003c/div\u003e \u003cdiv class=\"DefinitionListEntry\"\u003e \u003cdiv class=\"Term\"\u003eCI\u003c/div\u003e \u003cdiv class=\"Description\"\u003e \u003cp\u003econfidence interval\u003c/p\u003e \u003c/div\u003e \u003c/div\u003e \u003cdiv class=\"DefinitionListEntry\"\u003e \u003cdiv class=\"Term\"\u003eCRediT\u003c/div\u003e \u003cdiv class=\"Description\"\u003e \u003cp\u003eContributor Roles Taxonomy\u003c/p\u003e \u003c/div\u003e \u003c/div\u003e \u003cdiv class=\"DefinitionListEntry\"\u003e \u003cdiv class=\"Term\"\u003ecycle\u0026times;phase\u003c/div\u003e \u003cdiv class=\"Description\"\u003e \u003cp\u003egait cycle by fixed\u0026mdash;percentage cycle bin based on Perry timing references\u003c/p\u003e \u003c/div\u003e \u003c/div\u003e \u003cdiv class=\"DefinitionListEntry\"\u003e \u003cdiv class=\"Term\"\u003edb5\u003c/div\u003e \u003cdiv class=\"Description\"\u003e \u003cp\u003eDaubechies\u0026mdash;5 wavelet\u003c/p\u003e \u003c/div\u003e \u003c/div\u003e \u003cdiv class=\"DefinitionListEntry\"\u003e \u003cdiv class=\"Term\"\u003eDUA\u003c/div\u003e \u003cdiv class=\"Description\"\u003e \u003cp\u003eData Use Agreement\u003c/p\u003e \u003c/div\u003e \u003c/div\u003e \u003cdiv class=\"DefinitionListEntry\"\u003e \u003cdiv class=\"Term\"\u003eEMG\u003c/div\u003e \u003cdiv class=\"Description\"\u003e \u003cp\u003eelectromyography\u003c/p\u003e \u003c/div\u003e \u003c/div\u003e \u003cdiv class=\"DefinitionListEntry\"\u003e \u003cdiv class=\"Term\"\u003eF18\u003c/div\u003e \u003cdiv class=\"Description\"\u003e \u003cp\u003e18\u0026mdash;feature time/frequency/wavelet vector\u003c/p\u003e \u003c/div\u003e \u003c/div\u003e \u003cdiv class=\"DefinitionListEntry\"\u003e \u003cdiv class=\"Term\"\u003eF28\u003c/div\u003e \u003cdiv class=\"Description\"\u003e \u003cp\u003eF18\u0026thinsp;+\u0026thinsp;NL10\u003c/p\u003e \u003c/div\u003e \u003c/div\u003e \u003cdiv class=\"DefinitionListEntry\"\u003e \u003cdiv class=\"Term\"\u003eFD\u003c/div\u003e \u003cdiv class=\"Description\"\u003e \u003cp\u003efractal dimension\u003c/p\u003e \u003c/div\u003e \u003c/div\u003e \u003cdiv class=\"DefinitionListEntry\"\u003e \u003cdiv class=\"Term\"\u003eFT\u003c/div\u003e \u003cdiv class=\"Description\"\u003e \u003cp\u003efast\u0026mdash;fibre fraction\u003c/p\u003e \u003c/div\u003e \u003c/div\u003e \u003cdiv class=\"DefinitionListEntry\"\u003e \u003cdiv class=\"Term\"\u003eGM\u003c/div\u003e \u003cdiv class=\"Description\"\u003e \u003cp\u003egastrocnemius medialis\u003c/p\u003e \u003c/div\u003e \u003c/div\u003e \u003cdiv class=\"DefinitionListEntry\"\u003e \u003cdiv class=\"Term\"\u003eHS\u003c/div\u003e \u003cdiv class=\"Description\"\u003e \u003cp\u003eheel strike\u003c/p\u003e \u003c/div\u003e \u003c/div\u003e \u003cdiv class=\"DefinitionListEntry\"\u003e \u003cdiv class=\"Term\"\u003eIC\u003c/div\u003e \u003cdiv class=\"Description\"\u003e \u003cp\u003einitial contact\u003c/p\u003e \u003c/div\u003e \u003c/div\u003e \u003cdiv class=\"DefinitionListEntry\"\u003e \u003cdiv class=\"Term\"\u003eIITP\u003c/div\u003e \u003cdiv class=\"Description\"\u003e \u003cp\u003eInstitute of Information \u0026amp; Communications Technology Planning \u0026amp; Evaluation\u003c/p\u003e \u003c/div\u003e \u003c/div\u003e \u003cdiv class=\"DefinitionListEntry\"\u003e \u003cdiv class=\"Term\"\u003eIRB\u003c/div\u003e \u003cdiv class=\"Description\"\u003e \u003cp\u003eInstitutional Review Board\u003c/p\u003e \u003c/div\u003e \u003c/div\u003e \u003cdiv class=\"DefinitionListEntry\"\u003e \u003cdiv class=\"Term\"\u003eISw\u003c/div\u003e \u003cdiv class=\"Description\"\u003e \u003cp\u003einitial swing\u003c/p\u003e \u003c/div\u003e \u003c/div\u003e \u003cdiv class=\"DefinitionListEntry\"\u003e \u003cdiv class=\"Term\"\u003eJNER\u003c/div\u003e \u003cdiv class=\"Description\"\u003e \u003cp\u003eJournal of NeuroEngineering and Rehabilitation\u003c/p\u003e \u003c/div\u003e \u003c/div\u003e \u003cdiv class=\"DefinitionListEntry\"\u003e \u003cdiv class=\"Term\"\u003eLLM\u003c/div\u003e \u003cdiv class=\"Description\"\u003e \u003cp\u003elarge language model\u003c/p\u003e \u003c/div\u003e \u003c/div\u003e \u003cdiv class=\"DefinitionListEntry\"\u003e \u003cdiv class=\"Term\"\u003eLOSO\u003c/div\u003e \u003cdiv class=\"Description\"\u003e \u003cp\u003eleave\u0026mdash;one\u0026mdash;subject\u0026mdash;out\u003c/p\u003e \u003c/div\u003e \u003c/div\u003e \u003cdiv class=\"DefinitionListEntry\"\u003e \u003cdiv class=\"Term\"\u003eLOTO\u003c/div\u003e \u003cdiv class=\"Description\"\u003e \u003cp\u003eleave\u0026mdash;one\u0026mdash;protocol\u0026mdash;condition\u0026mdash;out\u003c/p\u003e \u003c/div\u003e \u003c/div\u003e \u003cdiv class=\"DefinitionListEntry\"\u003e \u003cdiv class=\"Term\"\u003eLR\u003c/div\u003e \u003cdiv class=\"Description\"\u003e \u003cp\u003eloading response\u003c/p\u003e \u003c/div\u003e \u003c/div\u003e \u003cdiv class=\"DefinitionListEntry\"\u003e \u003cdiv class=\"Term\"\u003eMAV\u003c/div\u003e \u003cdiv class=\"Description\"\u003e \u003cp\u003emean absolute value\u003c/p\u003e \u003c/div\u003e \u003c/div\u003e \u003cdiv class=\"DefinitionListEntry\"\u003e \u003cdiv class=\"Term\"\u003eMAD\u003c/div\u003e \u003cdiv class=\"Description\"\u003e \u003cp\u003emedian absolute deviation\u003c/p\u003e \u003c/div\u003e \u003c/div\u003e \u003cdiv class=\"DefinitionListEntry\"\u003e \u003cdiv class=\"Term\"\u003eMSt\u003c/div\u003e \u003cdiv class=\"Description\"\u003e \u003cp\u003emid stance\u003c/p\u003e \u003c/div\u003e \u003c/div\u003e \u003cdiv class=\"DefinitionListEntry\"\u003e \u003cdiv class=\"Term\"\u003eMSw\u003c/div\u003e \u003cdiv class=\"Description\"\u003e \u003cp\u003emid swing\u003c/p\u003e \u003c/div\u003e \u003c/div\u003e \u003cdiv class=\"DefinitionListEntry\"\u003e \u003cdiv class=\"Term\"\u003eMSIT\u003c/div\u003e \u003cdiv class=\"Description\"\u003e \u003cp\u003eMinistry of Science and ICT\u003c/p\u003e \u003c/div\u003e \u003c/div\u003e \u003cdiv class=\"DefinitionListEntry\"\u003e \u003cdiv class=\"Term\"\u003eNL10\u003c/div\u003e \u003cdiv class=\"Description\"\u003e \u003cp\u003e10\u0026mdash;feature nonlinear\u0026mdash;dynamics vector\u003c/p\u003e \u003c/div\u003e \u003c/div\u003e \u003cdiv class=\"DefinitionListEntry\"\u003e \u003cdiv class=\"Term\"\u003eNW\u003c/div\u003e \u003cdiv class=\"Description\"\u003e \u003cp\u003enot worn\u003c/p\u003e \u003c/div\u003e \u003c/div\u003e \u003cdiv class=\"DefinitionListEntry\"\u003e \u003cdiv class=\"Term\"\u003ePCSA\u003c/div\u003e \u003cdiv class=\"Description\"\u003e \u003cp\u003ephysiological cross\u0026mdash;sectional area\u003c/p\u003e \u003c/div\u003e \u003c/div\u003e \u003cdiv class=\"DefinitionListEntry\"\u003e \u003cdiv class=\"Term\"\u003ePE\u003c/div\u003e \u003cdiv class=\"Description\"\u003e \u003cp\u003eworn\u0026mdash;active suit\u003c/p\u003e \u003c/div\u003e \u003c/div\u003e \u003cdiv class=\"DefinitionListEntry\"\u003e \u003cdiv class=\"Term\"\u003ePSw\u003c/div\u003e \u003cdiv class=\"Description\"\u003e \u003cp\u003epre\u0026mdash;swing\u003c/p\u003e \u003c/div\u003e \u003c/div\u003e \u003cdiv class=\"DefinitionListEntry\"\u003e \u003cdiv class=\"Term\"\u003eRF\u003c/div\u003e \u003cdiv class=\"Description\"\u003e \u003cp\u003eRandom Forest (classifier) or rectus femoris (muscle, by context)\u003c/p\u003e \u003c/div\u003e \u003c/div\u003e \u003cdiv class=\"DefinitionListEntry\"\u003e \u003cdiv class=\"Term\"\u003eRMS\u003c/div\u003e \u003cdiv class=\"Description\"\u003e \u003cp\u003eroot mean square\u003c/p\u003e \u003c/div\u003e \u003c/div\u003e \u003cdiv class=\"DefinitionListEntry\"\u003e \u003cdiv class=\"Term\"\u003eRQA\u003c/div\u003e \u003cdiv class=\"Description\"\u003e \u003cp\u003erecurrence quantification analysis\u003c/p\u003e \u003c/div\u003e \u003c/div\u003e \u003cdiv class=\"DefinitionListEntry\"\u003e \u003cdiv class=\"Term\"\u003eSampEn\u003c/div\u003e \u003cdiv class=\"Description\"\u003e \u003cp\u003esample entropy\u003c/p\u003e \u003c/div\u003e \u003c/div\u003e \u003cdiv class=\"DefinitionListEntry\"\u003e \u003cdiv class=\"Term\"\u003eSD\u003c/div\u003e \u003cdiv class=\"Description\"\u003e \u003cp\u003estandard deviation\u003c/p\u003e \u003c/div\u003e \u003c/div\u003e \u003cdiv class=\"DefinitionListEntry\"\u003e \u003cdiv class=\"Term\"\u003eSENIAM\u003c/div\u003e \u003cdiv class=\"Description\"\u003e \u003cp\u003eSurface ElectroMyoGraphy for the Non\u0026mdash;Invasive Assessment of Muscles\u003c/p\u003e \u003c/div\u003e \u003c/div\u003e \u003cdiv class=\"DefinitionListEntry\"\u003e \u003cdiv class=\"Term\"\u003esEMG\u003c/div\u003e \u003cdiv class=\"Description\"\u003e \u003cp\u003esurface electromyography\u003c/p\u003e \u003c/div\u003e \u003c/div\u003e \u003cdiv class=\"DefinitionListEntry\"\u003e \u003cdiv class=\"Term\"\u003eSSC\u003c/div\u003e \u003cdiv class=\"Description\"\u003e \u003cp\u003eslope sign change\u003c/p\u003e \u003c/div\u003e \u003c/div\u003e \u003cdiv class=\"DefinitionListEntry\"\u003e \u003cdiv class=\"Term\"\u003eTA\u003c/div\u003e \u003cdiv class=\"Description\"\u003e \u003cp\u003etibialis anterior\u003c/p\u003e \u003c/div\u003e \u003c/div\u003e \u003cdiv class=\"DefinitionListEntry\"\u003e \u003cdiv class=\"Term\"\u003eTFL\u003c/div\u003e \u003cdiv class=\"Description\"\u003e \u003cp\u003etensor fasciae latae\u003c/p\u003e \u003c/div\u003e \u003c/div\u003e \u003cdiv class=\"DefinitionListEntry\"\u003e \u003cdiv class=\"Term\"\u003eTO\u003c/div\u003e \u003cdiv class=\"Description\"\u003e \u003cp\u003etoe\u0026mdash;off\u003c/p\u003e \u003c/div\u003e \u003c/div\u003e \u003cdiv class=\"DefinitionListEntry\"\u003e \u003cdiv class=\"Term\"\u003eTSt\u003c/div\u003e \u003cdiv class=\"Description\"\u003e \u003cp\u003eterminal stance\u003c/p\u003e \u003c/div\u003e \u003c/div\u003e \u003cdiv class=\"DefinitionListEntry\"\u003e \u003cdiv class=\"Term\"\u003eTSw\u003c/div\u003e \u003cdiv class=\"Description\"\u003e \u003cp\u003eterminal swing\u003c/p\u003e \u003c/div\u003e \u003c/div\u003e \u003cdiv class=\"DefinitionListEntry\"\u003e \u003cdiv class=\"Term\"\u003eUE\u003c/div\u003e \u003cdiv class=\"Description\"\u003e \u003cp\u003eworn\u0026mdash;inactive suit\u003c/p\u003e \u003c/div\u003e \u003c/div\u003e \u003cdiv class=\"DefinitionListEntry\"\u003e \u003cdiv class=\"Term\"\u003eWEF\u003c/div\u003e \u003cdiv class=\"Description\"\u003e \u003cp\u003ewavelet ensemble fusion\u003c/p\u003e \u003c/div\u003e \u003c/div\u003e \u003cdiv class=\"DefinitionListEntry\"\u003e \u003cdiv class=\"Term\"\u003eWL\u003c/div\u003e \u003cdiv class=\"Description\"\u003e \u003cp\u003ewaveform length\u003c/p\u003e \u003c/div\u003e \u003c/div\u003e \u003cdiv class=\"DefinitionListEntry\"\u003e \u003cdiv class=\"Term\"\u003eZCR\u003c/div\u003e \u003cdiv class=\"Description\"\u003e \u003cp\u003ezero\u0026mdash;crossing rate.\u003c/p\u003e \u003c/div\u003e \u003c/div\u003e \u003c/div\u003e"},{"header":"Declarations","content":"\u003cp\u003eTrial registration\u003cbr\u003e\u0026nbsp;Not applicable. This study is a recognition-stage methodological benchmark and does not evaluate prospectively assigned health-care interventions against clinical health outcomes.\u003c/p\u003e\n\u003ch3\u003eEthics approval and consent to participate\u003c/h3\u003e\n\u003cp\u003eThis study was approved by the Institutional Review Board (IRB) of the Electronics and Telecommunications Research Institute (ETRI) under approval number N01-202409-01-020. All participants were healthy adult men aged \u0026ge;19 years and provided written informed consent prior to enrollment, including consent for the experimental procedures described in Methods. All procedures were performed in accordance with the Declaration of Helsinki and the relevant guidelines and regulations.\u003c/p\u003e\n\u003ch3\u003eConsent for publication\u003c/h3\u003e\n\u003cp\u003eNot applicable. This manuscript contains no individually identifiable participant data, images, or videos; all figures and tables present only de-identified, aggregate, or schematic content.\u003c/p\u003e\n\u003ch3\u003eAvailability of data and materials\u003c/h3\u003e\n\u003cp\u003eThe openly reproducible component of this study is the processed-feature recognition benchmark. The Zenodo deposit [44] contains processed feature tensors, metadata, the official split manifest (splits/loso_manifest_v1.csv), official scoring script (metrics/score.py), H-MD-WEF + RF baseline source code, source values, reproduce_all.sh, and the reproducibility lock (seed = 42; macOS 26.3 arm64; Python 3.11.5; numpy 1.26.4; scipy 1.15.2; scikit-learn 1.7.1; commit 4e38a5b409e338c1e890bb64f86461fa4cd46926), which reproduce the reported Tables/Figures at the processed-feature level. Raw-signal preprocessing and EMG-derived HS/TO event extraction are documented for transparency but are not fully rerunnable from the public package because raw EMG redistribution is restricted by IRB-approved consent. During peer review, an anonymous reviewer-only Zenodo access link is provided through the JNER editorial system and converted to the public DOI on acceptance. Raw EMG may be requested for non-commercial academic research under a Data Use Agreement, subject to IRB/consent constraints.\u003c/p\u003e\n\u003ch3\u003eCompeting interests\u003c/h3\u003e\n\u003cp\u003eThe authors declare no competing interests.\u003c/p\u003e\n\u003ch3\u003eFunding\u003c/h3\u003e\n\u003cp\u003eThis work was supported by the Institute of Information \u0026amp; Communications Technology Planning \u0026amp; Evaluation (IITP), funded by the Ministry of Science and ICT (MSIT), Republic of Korea (Project No.\u0026nbsp;RS-2022-II220025, \u0026ldquo;Development of soft-suit technology to support human motor ability\u0026rdquo;). The funder had no role in study design; data collection, analysis, or interpretation; manuscript preparation; or the decision to publish.\u003c/p\u003e\n\u003ch3\u003eAuthors\u0026rsquo; contributions\u003c/h3\u003e\n\u003cp\u003eCRediT contributor roles (initials):\u003c/p\u003e\n\u003cul\u003e\n \u003cli\u003eSBL (Song-Bi Lee): Conceptualization; Methodology; Formal analysis; Data curation; Visualization; Writing \u0026ndash; original draft.\u003c/li\u003e\n \u003cli\u003eDWL (Dong-Woo Lee): Data curation; Funding acquisition; Project administration; Writing \u0026ndash; review \u0026amp; editing.\u003c/li\u003e\n \u003cli\u003eYK (Yongjun Kim): Data curation; Software; Validation; Writing \u0026ndash; review \u0026amp; editing.\u003c/li\u003e\n \u003cli\u003eGH (Gisu Heo): Writing \u0026ndash; review \u0026amp; editing.\u003c/li\u003e\n \u003cli\u003eCO (Changmok Oh): Writing \u0026ndash; review \u0026amp; editing.\u003c/li\u003e\n \u003cli\u003eSE (Suyeong Eom): Data curation; IRB protocol development.\u003c/li\u003e\n\u003c/ul\u003e\n\u003cp\u003eAll authors read and approved the final manuscript.\u003c/p\u003e\n\u003ch3\u003eAcknowledgements\u003c/h3\u003e\n\u003cp\u003eWe thank all participants. We used Elsevier Language Editing Services for English-language editing of this manuscript.\u003c/p\u003e\n\u003ch3\u003eUse of large language models\u003c/h3\u003e\n\u003cp\u003eSee Methods: Use of AI tools.\u003c/p\u003e\n\u003ch3\u003eReporting guideline checklist\u003c/h3\u003e\n\u003cp\u003eA completed reporting guideline checklist (STROBE, with item-to-manuscript mapping and N/A justifications where appropriate) is included as a separate submission file for peer review. Because the present study is a machine-learning prediction-style recognition benchmark rather than a clinical observational study, we additionally provide a TRIPOD+AI-style item-mapping supplement that addresses model reporting items that are not fully covered by STROBE (data split definition, leakage prevention, calibration regimes, missing/filtered samples, model hyperparameters, performance uncertainty, and the absence of external validation). The TRIPOD+AI mapping is intended only as supplementary transparency, not as a claim that this work is a clinical prediction model.\u003c/p\u003e"},{"header":"References","content":"\u003col\u003e\n\u003cli\u003eSawicki, G. S. et al. The exoskeleton expansion: improving walking and running economy. J. NeuroEngineering Rehabil. 17, 25 (2020).\u003c/li\u003e\n\u003cli\u003eYoung, A. J. \u0026amp; Ferris, D. P. State of the art and future directions for lower limb robotic exoskeletons. IEEE Trans. Neural Syst. Rehabil. Eng. 25, 171\u0026ndash;182 (2017).\u003c/li\u003e\n\u003cli\u003ePanizzolo, F. A. et al. A biologically-inspired multi-joint soft exosuit that can reduce the energy cost of loaded walking. J. NeuroEngineering Rehabil. 13, 43 (2016).\u003c/li\u003e\n\u003cli\u003eDing, Y. et al. Human-in-the-loop optimization of hip assistance with a soft exosuit during walking. Sci. Robot. 3, eaar5438 (2018).\u003c/li\u003e\n\u003cli\u003eZhang, J. et al. Human-in-the-loop optimization of exoskeleton assistance during walking. Science 356, 1280\u0026ndash;1284 (2017).\u003c/li\u003e\n\u003cli\u003eMalcolm, P. et al. A simple exoskeleton that assists plantarflexion can reduce the metabolic cost of human walking. PLoS ONE 8, e56137 (2013).\u003c/li\u003e\n\u003cli\u003eQuinlivan, B. T. et al. Assistance magnitude versus metabolic cost reductions for a tethered multiarticular soft exosuit. Sci. Robot. 2, eaah4416 (2017).\u003c/li\u003e\n\u003cli\u003eHuang, H. et al. Continuous locomotion-mode identification for prosthetic legs based on neuromuscular-mechanical fusion. IEEE Trans. Biomed. Eng. 58, 2867\u0026ndash;2875 (2011).\u003c/li\u003e\n\u003cli\u003ePhinyomark, A. et al. Feature reduction and selection for EMG signal classification. Expert Syst. Appl. 39, 7420\u0026ndash;7431 (2012).\u003c/li\u003e\n\u003cli\u003eEnglehart, K. \u0026amp; Hudgins, B. A robust, real-time control scheme for multifunction myoelectric control. IEEE Trans. Biomed. Eng. 50, 848\u0026ndash;854 (2003).\u003c/li\u003e\n\u003cli\u003eScheme, E. \u0026amp; Englehart, K. Electromyogram pattern recognition for control of powered upper-limb prostheses: state of the art and challenges for clinical use. J. Rehabil. Res. Dev. 48, 643\u0026ndash;659 (2011).\u003c/li\u003e\n\u003cli\u003eFarina, D. et al. The extraction of neural information from the surface EMG for the control of upper-limb prostheses: emerging avenues and challenges. IEEE Trans. Neural Syst. Rehabil. Eng. 22, 797\u0026ndash;809 (2014).\u003c/li\u003e\n\u003cli\u003eHargrove, L. J. et al. Myoelectric pattern recognition outperforms direct control for transhumeral amputees with targeted muscle reinnervation: a randomized clinical trial. Sci. Rep. 7, 13840 (2017).\u003c/li\u003e\n\u003cli\u003eLiu, J. et al. EMG-based real-time linear-nonlinear cascade regression decoding of shoulder, elbow and wrist movements in able-bodied persons and stroke survivors. IEEE Trans. Biomed. Eng. 67, 1272\u0026ndash;1281 (2020).\u003c/li\u003e\n\u003cli\u003eNazarpour, K. et al. A note on the probability distribution function of the surface electromyogram signal. Brain Res. Bull. 90, 73\u0026ndash;79 (2013).\u003c/li\u003e\n\u003cli\u003eDe Luca, C. J. The use of surface electromyography in biomechanics. J. Appl. Biomech. 13, 135\u0026ndash;163 (1997).\u003c/li\u003e\n\u003cli\u003eMerletti, R. \u0026amp; Farina, D. Analysis of intramuscular electromyogram signals. Phil. Trans. R. Soc. A 367, 357\u0026ndash;368 (2009).\u003c/li\u003e\n\u003cli\u003eClancy, E. A. et al. Sampling, noise-reduction and amplitude estimation issues in surface electromyography. J. Electromyogr. Kinesiol. 12, 1\u0026ndash;16 (2002).\u003c/li\u003e\n\u003cli\u003ePhinyomark, A. et al. EMG feature evaluation for improving myoelectric pattern recognition robustness. Expert Syst. Appl. 40, 4832\u0026ndash;4840 (2013).\u003c/li\u003e\n\u003cli\u003eTkach, D. et al. Study of stability of time-domain features for electromyographic pattern recognition. J. NeuroEngineering Rehabil. 7, 21 (2010).\u003c/li\u003e\n\u003cli\u003eHudgins, B. et al. A new strategy for multifunction myoelectric control. IEEE Trans. Biomed. Eng. 40, 82\u0026ndash;94 (1993).\u003c/li\u003e\n\u003cli\u003eGraupe, D. \u0026amp; Cline, W. K. Functional separation of EMG signals via ARMA identification methods for prosthesis control purposes. IEEE Trans. Syst. Man Cybern. 5, 252\u0026ndash;259 (1975).\u003c/li\u003e\n\u003cli\u003eChan, A. D. C. \u0026amp; Englehart, K. B. Continuous myoelectric control for powered prostheses using hidden Markov models. IEEE Trans. Biomed. Eng. 52, 121\u0026ndash;124 (2005).\u003c/li\u003e\n\u003cli\u003eMallat, S. G. A theory for multiresolution signal decomposition: the wavelet representation. IEEE Trans. Pattern Anal. Mach. Intell. 11, 674\u0026ndash;693 (1989).\u003c/li\u003e\n\u003cli\u003eDaubechies, I. Ten Lectures on Wavelets (SIAM, 1992).\u003c/li\u003e\n\u003cli\u003eEnglehart, K. et al. A wavelet-based continuous classification scheme for multifunction myoelectric control. IEEE Trans. Biomed. Eng. 48, 302\u0026ndash;311 (2001).\u003c/li\u003e\n\u003cli\u003eRichman, J. S. \u0026amp; Moorman, J. R. Physiological time-series analysis using approximate entropy and sample entropy. Am. J. Physiol. Heart Circ. Physiol. 278, H2039\u0026ndash;H2049 (2000).\u003c/li\u003e\n\u003cli\u003eHiguchi, T. Approach to an irregular time series on the basis of the fractal theory. Physica D 31, 277\u0026ndash;283 (1988).\u003c/li\u003e\n\u003cli\u003eHausdorff, J. M. Gait dynamics in Parkinson\u0026apos;s disease: common and distinct behavior among stride length, gait variability, and fractal-like scaling. Chaos 19, 026113 (2009).\u003c/li\u003e\n\u003cli\u003eHausdorff, J. M. et al. Gait variability and fall risk in community-living older adults: a 1-year prospective study. Arch. Phys. Med. Rehabil. 82, 1050\u0026ndash;1056 (2001).\u003c/li\u003e\n\u003cli\u003eRosenstein, M. T., Collins, J. J. \u0026amp; De Luca, C. J. A practical method for calculating largest Lyapunov exponents from small data sets. Physica D 65, 117\u0026ndash;134 (1993).\u003c/li\u003e\n\u003cli\u003eKatz, M. J. Fractals and the analysis of waveforms. Comput. Biol. Med. 18, 145\u0026ndash;156 (1988).\u003c/li\u003e\n\u003cli\u003ePincus, S. M. Approximate entropy as a measure of system complexity. Proc. Natl. Acad. Sci. USA 88, 2297\u0026ndash;2301 (1991).\u003c/li\u003e\n\u003cli\u003eBandt, C. \u0026amp; Pompe, B. Permutation entropy: a natural complexity measure for time series. Phys. Rev. Lett. 88, 174102 (2002).\u003c/li\u003e\n\u003cli\u003eHurst, H. E. Long-term storage capacity of reservoirs. Trans. Am. Soc. Civ. Eng. 116, 770\u0026ndash;799 (1951).\u003c/li\u003e\n\u003cli\u003eWebber, C. L. Jr. \u0026amp; Zbilut, J. P. Dynamical assessment of physiological systems and states using recurrence plot strategies. J. Appl. Physiol. 76, 965\u0026ndash;973 (1994).\u003c/li\u003e\n\u003cli\u003eMarwan, N. et al. Recurrence plots for the analysis of complex systems. Phys. Rep. 438, 237\u0026ndash;329 (2007).\u003c/li\u003e\n\u003cli\u003eHermens, H. J. et al. Development of recommendations for SEMG sensors and sensor placement procedures. J. Electromyogr. Kinesiol. 10, 361\u0026ndash;374 (2000).\u003c/li\u003e\n\u003cli\u003ePerry, J. \u0026amp; Burnfield, J. M. Gait Analysis: Normal and Pathological Function 2nd edn (SLACK, 2010).\u003c/li\u003e\n\u003cli\u003eWard, S. R. et al. Are current measurements of lower extremity muscle architecture accurate? Clin. Orthop. Relat. Res. 467, 1074\u0026ndash;1082 (2009).\u003c/li\u003e\n\u003cli\u003eJohnson, M. A., Polgar, J., Weightman, D. \u0026amp; Appleton, D. Data on the distribution of fibre types in thirty-six human muscles: an autopsy study. J. Neurol. Sci. 18, 111\u0026ndash;129 (1973).\u003c/li\u003e\n\u003cli\u003ePedregosa, F. et al. Scikit-learn: machine learning in Python. J. Mach. Learn. Res. 12, 2825\u0026ndash;2830 (2011).\u003c/li\u003e\n\u003cli\u003eBreiman, L. Random forests. Mach. Learn. 45, 5\u0026ndash;32 (2001).\u003c/li\u003e\n\u003cli\u003eLee, S.-B. et al. MD-WEF recognition-stage benchmark v1.0 (processed-feature tensors, split manifest, scoring scripts, and baseline code) [dataset]. Zenodo. https://doi.org/10.5281/zenodo.19547468.\u003c/li\u003e\n\u003c/ol\u003e"}],"fulltextSource":"","fullText":"","funders":[],"hasAdminPriorityOnWorkflow":false,"hasManuscriptDocX":true,"hasOptedInToPreprint":true,"hasPassedJournalQc":"","hasAnyPriority":false,"hideJournal":false,"highlight":"","institution":"","isAcceptedByJournal":false,"isAuthorSuppliedPdf":false,"isDeskRejected":"","isHiddenFromSearch":false,"isInQc":false,"isInWorkflow":false,"isPdf":false,"isPdfUpToDate":true,"isWithdrawnOrRetracted":false,"journal":{"display":true,"email":"[email protected]","identity":"journal-of-neuroengineering-and-rehabilitation","isNatureJournal":false,"hasQc":true,"allowDirectSubmit":false,"externalIdentity":"jner","sideBox":"Learn more about [Journal of NeuroEngineering and Rehabilitation](http://jneuroengrehab.biomedcentral.com/)","snPcode":"12984","submissionUrl":"https://submission.nature.com/new-submission/12984/3","title":"Journal of NeuroEngineering and Rehabilitation","twitterHandle":"@BioMedCentral","acdcEnabled":true,"dfaEnabled":true,"editorialSystem":"em","reportingPortfolio":"BMC/SO AJ","inReviewEnabled":true,"inReviewRevisionsEnabled":true},"keywords":"Cable-driven suit, surface electromyography, protocol-state discriminability, subject-independent benchmark, leave-one-subject-out, calibration, computational latency, reproducibility","lastPublishedDoi":"10.21203/rs.3.rs-9590227/v1","lastPublishedDoiUrl":"https://doi.org/10.21203/rs.3.rs-9590227/v1","license":{"name":"CC BY 4.0","url":"https://creativecommons.org/licenses/by/4.0/"},"manuscriptAbstract":"\u003cp\u003e\u003cb\u003eBackground\u003c/b\u003e\u003c/p\u003e \u003cp\u003eWe present a null-referenced processed-feature leave-one-subject-out (LOSO) transfer-boundary benchmark for surface electromyography (sEMG) protocol-state discriminability in healthy adult men using a cable-driven walking-assistance suit. Protocol states are defined as not-worn (NW), worn-inactive suit (UE), and worn-active suit (PE). PE denotes active assistance in the same worn hardware configuration as UE; actuator commands, cable-tension traces, and torque profiles were excluded from predictors.\u003c/p\u003e\u003cp\u003e\u003cb\u003eMethods\u003c/b\u003e\u003c/p\u003e \u003cp\u003eData from 10 healthy adult men were analyzed. The processed dataset comprised 29,993 EMG-derived gait cycles (239,944 cycle\u0026times;phase samples; 10-channel sEMG at 2,148 Hz). We defined a fixed subject-independent LOSO benchmark with a balanced manifest capped at \u0026le;\u0026thinsp;50 cycles per recording file, cycle-level macro-F1 as the primary endpoint, separated calibration regimes, and offline computational-cost readouts. Leakage control denotes subject-level split separation and exclusion of held-out-subject labels and normalization statistics, not removal of residual order, fatigue, electrode-state, or protocol-speed confounding. The reproducible unit is the processed-feature recognition benchmark; raw EMG preprocessing and EMG-derived event extraction are documented but not fully rerunnable from the public package because raw time series are available only under DUA.\u003c/p\u003e\u003cp\u003e\u003cb\u003eResults\u003c/b\u003e\u003c/p\u003e \u003cp\u003eCycle-level macro-F1 was 0.41\u0026ndash;0.43 within condition and 0.18\u0026ndash;0.41 across protocol-condition transfer cells. Direct cycle-level Random Forest and multinomial logistic-regression comparators preserved the same diagonal-versus-off-diagonal ordering. Offline computational-cost readouts separated a high-cost reference configuration from a lower-latency comparator (inference-only mean 1.36 ms; lower-latency end-to-end mean 115 ms; high-cost end-to-end mean 2,697 ms). The lower-latency comparator was not presented as a deployment-ready control claim. Uncertainty was estimated at the held-out-subject level; cycle and phase counts should not be interpreted as independent participant counts.\u003c/p\u003e\u003cp\u003e\u003cb\u003eConclusions\u003c/b\u003e\u003c/p\u003e \u003cp\u003eThe contribution is a fixed benchmark procedure\u0026mdash;split manifest, transfer grid, null references, scoring endpoint, and reproducibility package\u0026mdash;rather than a new classifier or deployment controller. The modest macro-F1 values are part of the benchmark finding: under fixed leakage-controlled LOSO scoring, processed-feature EMG-only recognition showed limited subject-independent transfer across protocol-condition bundles. Findings should be interpreted as protocol-condition transfer across bundled terrain/platform/speed conditions in this healthy adult male proof-of-concept cohort.\u003c/p\u003e\u003cp\u003e\u003cb\u003eTrial registration\u003c/b\u003e\u003c/p\u003e \u003cp\u003eNot applicable. This study is a recognition-stage methodological benchmark and does not evaluate prospectively assigned health-care interventions against clinical health outcomes.\u003c/p\u003e\u003cp\u003e\u003cb\u003eTrial registration\u003c/b\u003e\u003c/p\u003e \u003cp\u003eNot applicable.\u003c/p\u003e","manuscriptTitle":"A reproducible processed-feature LOSO benchmark for sEMG protocol-state discriminability in a cable-driven walking-assistance suit","msid":"","msnumber":"","nonDraftVersions":[{"code":1,"date":"2026-05-13 09:25:42","doi":"10.21203/rs.3.rs-9590227/v1","editorialEvents":[{"type":"communityComments","content":0},{"type":"reviewerAgreed","content":"260694246202246499566340651027369862437","date":"2026-05-12T02:49:03+00:00","index":"hide","fulltext":""},{"type":"reviewerAgreed","content":"228023843376371149270407628577814705659","date":"2026-05-09T04:41:09+00:00","index":"hide","fulltext":""},{"type":"reviewerAgreed","content":"235163886616472743168188670968712967618","date":"2026-05-08T13:58:30+00:00","index":"hide","fulltext":""},{"type":"reviewerAgreed","content":"7507209186550007936725795386399776254","date":"2026-05-07T07:51:15+00:00","index":"hide","fulltext":""},{"type":"reviewersInvited","content":"","date":"2026-05-05T03:28:47+00:00","index":"","fulltext":""},{"type":"editorAssigned","content":"","date":"2026-05-04T13:23:11+00:00","index":"","fulltext":""},{"type":"checksComplete","content":"","date":"2026-05-04T11:55:27+00:00","index":"","fulltext":""},{"type":"submitted","content":"Journal of NeuroEngineering and Rehabilitation","date":"2026-05-02T02:54:34+00:00","index":"","fulltext":""}],"status":"published","journal":{"display":true,"email":"[email protected]","identity":"journal-of-neuroengineering-and-rehabilitation","isNatureJournal":false,"hasQc":true,"allowDirectSubmit":false,"externalIdentity":"jner","sideBox":"Learn more about [Journal of NeuroEngineering and Rehabilitation](http://jneuroengrehab.biomedcentral.com/)","snPcode":"12984","submissionUrl":"https://submission.nature.com/new-submission/12984/3","title":"Journal of NeuroEngineering and Rehabilitation","twitterHandle":"@BioMedCentral","acdcEnabled":true,"dfaEnabled":true,"editorialSystem":"em","reportingPortfolio":"BMC/SO AJ","inReviewEnabled":true,"inReviewRevisionsEnabled":true}}],"origin":"","ownerIdentity":"5a4de0e8-8691-4d18-b10c-9f3e7ee9e8e2","owner":[],"postedDate":"May 13th, 2026","published":true,"recentEditorialEvents":[{"type":"reviewerAgreed","content":"260694246202246499566340651027369862437","date":"2026-05-12T02:49:03+00:00","index":31,"fulltext":""},{"type":"reviewerAgreed","content":"228023843376371149270407628577814705659","date":"2026-05-09T04:41:09+00:00","index":28,"fulltext":""},{"type":"reviewerAgreed","content":"235163886616472743168188670968712967618","date":"2026-05-08T13:58:30+00:00","index":27,"fulltext":""},{"type":"reviewerAgreed","content":"7507209186550007936725795386399776254","date":"2026-05-07T07:51:15+00:00","index":25,"fulltext":""},{"type":"reviewersInvited","content":"16","date":"2026-05-05T03:28:47+00:00","index":"","fulltext":""},{"type":"editorAssigned","content":"","date":"2026-05-04T13:23:11+00:00","index":"","fulltext":""},{"type":"checksComplete","content":"","date":"2026-05-04T11:55:27+00:00","index":"","fulltext":""},{"type":"submitted","content":"Journal of NeuroEngineering and Rehabilitation","date":"2026-05-02T02:54:34+00:00","index":"","fulltext":""}],"rejectedJournal":[],"revision":"","amendment":"","status":"under-review","subjectAreas":[],"tags":[],"updatedAt":"2026-05-13T09:25:43+00:00","versionOfRecord":[],"versionCreatedAt":"2026-05-13 09:25:42","video":"","vorDoi":"","vorDoiUrl":"","workflowStages":[]},"version":"v1","identity":"rs-9590227","journalConfig":"researchsquare"},"__N_SSP":true},"page":"/article/[identity]/[[...version]]","query":{"redirect":"/article/rs-9590227","identity":"rs-9590227","version":["v1"]},"buildId":"XKTyCvWXoU3ODBz1xrDgd","isFallback":false,"isExperimentalCompile":false,"dynamicIds":[84888],"gssp":true,"scriptLoader":[]}