Fatigue Severity Hidden by Joint Compensation: Sign-Constrained Ordinal Learning from Wearables

When a muscle group tires, the body often keeps the overall movement looking normal by shifting the work to other joints. After quadriceps fatigue, basketball players kept the total landing force unchanged because the ankle took over load that the knee and hip could no longer absorb (Impact of Quadriceps Muscle Fatigue on Ankle Joint Compensation Strategies During Single-Leg Vertical Jump Landing). This hiding is exactly what makes fatigue dangerous, and it makes fatigue hard to detect from wearable sensors. The severity of fatigue does not show up in any single quantity. It shows up in a signed redistribution: some joint quantities fall and others rise. We propose to teach a wearable-signal model an ordinal severity axis, meaning an axis whose values have a meaningful order but no fixed spacing. During training, the axis is required to move each laboratory-measured joint quantity in the direction biomechanics says it should. At test time only the wearable signal is needed.

In short, the model reads fatigue severity from muscle-activity sensors. It is trained so that as its severity estimate rises, its predicted knee moment and stiffness must fall and its predicted ankle power and sideways body sway must rise. This sign pattern comes from motion-capture studies and acts as a structural prior that ordinary ordinal classifiers do not have. The question is whether this prior, supplied only at training time, gives better and more trustworthy severity estimates from wearables than ordinal losses that know nothing about joint physics.

Core Claim

The postural–joint trade-off literature gives a consistent sign pattern for progressive quadriceps fatigue. In badminton scissor-jump landings, knee extension moment decreased and knee adduction moment, knee joint reaction force and coronal center-of-mass displacement increased (Analysis of Quadriceps Fatigue Effects on Lower Extremity Injury Risks During Landing Phases in Badminton Scissor Jump). In single-leg landings, hip and knee stiffness dropped while ankle plantarflexion power rose (Impact of Quadriceps Muscle Fatigue on Ankle Joint Compensation Strategies During Single-Leg Vertical Jump Landing). Those studies only contrast pre-fatigue with post-fatigue. We treat their sign pattern as a monotonicity hypothesis across the whole fatiguing sequence.

Our claim is that imposing this signed monotonicity between a learned severity scalar and privileged joint quantities makes the representation more ordered than ordinal losses alone. Here "privileged" means quantities available during training but not at test time, in this case OpenSim joint moments, stiffness and powers. We expect the benefit to be largest for subjects whose global outcome hides their fatigue. A shuffled-sign control should destroy the benefit. If it does not, the gain comes from generic multi-task supervision rather than from the trade-off structure, and the claim fails.

This differs from the only prior work combining movement signals with ordinal distribution learning, Ordinal Distribution Regression for Gait-based Age Estimation. That paper applies an Earth Mover's loss over binary age heads to video silhouettes split by body segment. It never uses joint kinetics, and it imposes no constraint linking the ordinal output to physically signed segment quantities.

Research Program

We would build a recording protocol modeled on the two landing studies. Each participant performs repeated single-leg landings while a quadriceps fatigue protocol runs to task failure. Every landing is recorded, not just the first and last. Each landing yields 8-channel surface electromyography, the wearable-grade input, together with laboratory motion capture, force plates and OpenSim joint quantities, which serve as the privileged training signal. The file does not say whether the original badminton or basketball recordings are released. We therefore plan new collection of roughly 30 to 40 participants and will request the existing data in parallel.

The target labels combine two of the input directions. The ordinal label is a fatigue stage derived from each landing's position in the sequence, expressed as a fraction of repetitions to failure. Stage boundaries are crude and progression is continuous, so we do not use one-hot stages. Instead we use label distribution learning, which trains against a probability spread over neighbouring stages (Label Distribution Learning). We spread mass with a discretized Beta profile centred on the normalized repetition index, so the profile can be skewed near task failure.

We train with the cumulative-distribution loss from Ordinal Label Distribution Learning. That paper showed that standard KL divergence is blind to order and replaced it with distances over cumulative distributions. We reuse its method, not its domain.

The encoder maps an electromyography window $x$ to an embedding $z$ and a scalar severity $s(z)$. On top of $s$ we add two pieces.

The first is an ordinal contrastive term. It is Rank-N-Contrast from Supervised Contrastive Regression, which orders embedding distances by label distance within a batch. We apply it only within each subject, since baseline muscle activity differs across people far more than fatigue changes it.

The second, and the new part, is a sign-constrained joint decoder. For each privileged joint quantity $k$, the predicted value is

$$\hat{j}_k = b_k(u) + \sigma_k\, g_k(s),$$

where $u$ is a per-subject baseline embedding taken from the first landings, $b_k$ is an unconstrained network, $\sigma_k \in \{-1,+1\}$ is the sign taken from the biomechanics literature, and $g_k$ is a monotone increasing network with positive weights. Because each $g_k$ is monotone by construction, the decoded joint curves can never cross the literature direction. This is the same structural guarantee that non-crossing ordinal methods enforce on decision thresholds (Noncrossing Ordinal Classification), here applied to coupled physical outputs instead of a single scalar target.

The training objective is

$$L = L_{\mathrm{CAD}}(p(x), d(x)) + \alpha\, L_{\mathrm{RNC}}(z) + \beta \sum_k (\hat{j}_k - j_k)^2,$$

where $p(x)$ is the predicted stage distribution, $d(x)$ is the soft Beta target, $L_{\mathrm{CAD}}$ is the cumulative absolute distance, $L_{\mathrm{RNC}}$ is the within-subject Rank-N-Contrast loss, and $j_k$ is the OpenSim value. Gradients from the joint-reconstruction term can only shape $s$ in directions consistent with the trade-off.

We will also test the unimodality assumption directly. Unimodal-Concentrated Loss: Fully Adaptive Label Distribution Learning for Ordinal Regression forces every predicted distribution to have a single peak. Motor-control work suggests the nervous system can switch between discrete strategies under stress (Effects of Perturbations to Balance on Neuromechanics of Fast Changes in Direction during Locomotion). Some subjects may therefore appear genuinely ambiguous between stages, for example by alternating knee-dominant and ankle-dominant landings. We will compare unimodal-constrained and unconstrained output heads on those subjects. Our model is not built to resolve this tension. The comparison is a falsifiable side question.

Baselines

The naive reference point is a standard cross-entropy stage classifier on the same encoder. A second reference is a direct scalar regressor on the normalized repetition index with an L1 loss. Improving Deep Regression with Tightness argues that this kind of regressor leaves representations loose. On AgeDB-DIR it reports MAE of 7.80 for plain regression versus 7.28 with its tightness regularizers. We include its multi-target and optimal-transport tightness regularizer as an additional ordinal-geometry baseline.

The strongest existing system for wearable fatigue staging is FES-sEMGNet from Recognition of Human Lower Limb Motion and Muscle Fatigue Status Using a Wearable FES-sEMG System. It is a CNN-LSTM over Hilbert-Huang time-frequency images and reports 89.02% accuracy on three fatigue states with 5 subjects. That number comes from a different, tiny dataset with an 8:1:1 split that is not stated to be subject-independent. It is not a target we can compare against directly, and no reported result was found on Lacuna for any method on progressive multi-stage landing fatigue with leave-subject-out evaluation.

The strongest ordinal-learning competitors we will run are:

All methods share the same encoder architecture, input windows, leave-one-subject-out folds and data augmentation. Each receives the same hyperparameter search budget of 40 random configurations, tuned on inner subject folds.

Two ablations isolate the key ingredient:

  • Unsigned decoder. The joint decoder is kept but the $\sigma_k$ signs and the monotonicity are removed, which gives plain multi-task use of the privileged signal.
  • Shuffled signs. The decoder is kept but the $\sigma_k$ are randomly permuted.

Beating the field means two things. Our model must reduce stage MAE and the cumulative-distribution error relative to the best ordinal competitor, with a paired subject-level test that does not overlap zero. It must also beat the unsigned-decoder ablation by a margin larger than the across-seed spread.

Why not just…?

Why not just use classic electromyography fatigue features with ordinal logistic regression? Falling median frequency and rising amplitude are well-known fatigue markers available from wearables. With roughly 30 subjects, a linear ordinal model could match a deep network, so this is a real threat. We will run it with the same folds. If it ties our model overall, the remaining contribution reduces to the compensation-masked subgroup. We would then report the method as a way to learn subject-specific redistribution patterns rather than as a better general fatigue detector.

Why not just add the joint quantities as auxiliary regression targets? This is the unsigned-decoder ablation. Privileged multi-task learning alone may supply most of the gain. If the signed model does not beat the unsigned one, and the shuffled-sign model is not clearly worse, the postural trade-off adds nothing and the core claim is refuted.

Why not just apply Rank-N-Contrast or CORE on the repetition index? CORE: Learning Consistent Ordinal REpresentations for Image Ordinal Estimation and Rank-N-Contrast produce ordered one-dimensional embeddings on image data; CORE cut MORPH MAE from 2.41 to 2.13. Neither ties the axis to physical directions. An axis ordered by repetition count could therefore track something other than fatigue, such as electrode drift or sweat, which also accumulates over trials. The test is whether the decoded joint quantities generalize to held-out subjects. A pure contrastive model has no joint decoder, so we fit a post-hoc linear probe from its embedding to the joint quantities and compare the probe's held-out accuracy against our decoder.

Proposed Experiments

The main evaluation is leave-one-subject-out on the new landing dataset. The metrics are:

  • stage MAE and exact-stage accuracy;
  • cumulative absolute distance between predicted and soft target distributions;
  • Spearman correlation between predicted severity and repetition index within each subject;
  • monotonicity violation rate, meaning the fraction of within-subject landing pairs whose predicted severity order contradicts their repetition order;
  • held-out reconstruction error of the OpenSim joint quantities.

The first decisive check is the masked subgroup. We identify subjects whose total landing impulse stays flat across the sequence, the compensation pattern reported for single-leg landings. We expect the gap between our model and the best ordinal competitor to be largest there. If it is not, the mechanism we claim is not what drives any gain.

The second decisive check is the sign-shuffle control. It should perform no better than the unsigned decoder.

The third decisive check is external. Without retraining, we apply the encoder's severity output to the three-state FES-sEMG recordings using the channel subset shared between the two setups. We test whether the predicted severity ranks no-fatigue below medium below extreme fatigue.

Finally, as a check that the soft-label and contrastive machinery is not broken on large gait data, we train the ordinal part of the model, without the joint decoder, on OULP-Age. The reference is ODR-GLCNN's reported MAE of 5.12 years (Ordinal Distribution Regression for Gait-based Age Estimation). We do not claim gains there.