Is the Crutch Really Carrying Weight? Detecting Simulated Impairment from Arm–Leg Load Timing

When a person truly cannot put full weight on a leg, their arms take over part of the job. The timing and size of the force pushed through a crutch or cane are then tied to what the injured leg is doing, stride by stride, by plain mechanics. A person who is only acting impaired can copy how an injured gait looks, including a shortened step or a guarded knee. What they have to build on purpose is the tight, unconscious timing between arm loading and leg contact. This proposal claims that this cross-limb timing structure, measured with an instrumented crutch and shin-worn motion sensors, separates genuine from simulated impairment. It also claims the separation grows over hours of unsupervised daily life, because keeping up a performance costs attention and attention runs out.

The idea borrows its central move from text-based deception detection. There, the grammatical shape of sentences, with every word removed, still reveals fabricated reviews (Syntactic Stylometry for Deception Detection). We treat the ordered sequence of arm and leg contact events as the motor counterpart of that word-free grammar. The first uses are practical and fairly low-stakes. One is checking that able-bodied research participants actually comply when told to simulate an impairment, a verification step that current transfer studies lack (Development of custom measurement system for biomechanical evaluation of independent wheelchair transfers). The other is telling patients who are prescribed partial weight-bearing when they have drifted away from their prescribed pattern. Using this for malingering in insurance or legal settings is a later question. We treat it as out of scope unless false-positive rates can be shown to be very low.

Core Claim

Separating genuine from voluntarily simulated impairment works better on the structure of arm–leg coupling than on the size of any single signal. Structure here means the order of events, the lags between them, and how stable those lags are from one stride to the next. Examples of such lags are how long after crutch strike the affected-side foot lands, and how peak crutch load lines up with that foot's stance phase. Surface magnitudes such as crutch peak force, stride length and joint angles are what a simulator watches and copies. The lags are not, because in genuine offloading they come out of load-sharing mechanics rather than from choice. Our second claim is ecological. In a single supervised lab session a simulator can hold their attention on the performance. Across days at home, with interruptions, distractions and fatigue, the imposed pattern should come apart first in exactly these coupling features. So a detector should become more accurate outside the lab, which is the reverse of how most wearable gait tools behave.

What Each Direction Contributes

The upper-extremity loading literature supplies the physical premise and the measuring tools. Custom six-axis load beams show that sitting-pivot transfers put vertical hand forces above 29% of body weight through both hands (Development of custom measurement system…). Clinician-rated transfer technique predicts joint moments and how fast they rise at the shoulder and elbow, with regression $R^2$ between 0.27 and 0.79 (The Relationship between Independent Transfer Skills and Upper Limb Kinetics in Wheelchair Users). Together these papers establish that arm loading is lawfully tied to how the rest of the body moves. Work on hand-carried loads during stair climbing shows the reverse link: arm loading shifts lower-limb joint stress and centre-of-pressure travel (Biomechanical effects of loading methods on the patellofemoral joint during stair climbing). All of this work is confined to the lab and bolted to fixed hardware. None of it asks whether the coupling can be faked. That is the gap we fill, using a load cell carried inside a crutch shaft instead of a force plate.

Stylometric deception detection supplies the analytic lens. Bag-of-words classifiers reach 89.8% on fabricated hotel reviews, well above human judges at 53.1–61.9% (Finding Deceptive Opinion Spam by Any Stretch of the Imagination). The more relevant result for us is that unlexicalized parse-tree rules alone reach 78.5% (Syntactic Stylometry for Deception Detection). That means structure carries signal independent of content, and it stays more stable across topics than vocabulary does. Behavioral stylometry in chess shows that style survives even in non-linguistic action sequences (Detecting Individual Decision-Making Style). We take the "content-free structure" idea from the first line and the "action sequences have style" idea from the second. We apply them to a question neither paper asks: whether a motor pattern is produced by constraint or by choice.

Ecological gait analysis supplies the setting and much of the signal processing. Deep networks now detect foot contacts in real-life walking bouts across six clinical groups, reaching 98% recall and 96% precision for initial contacts (Ecological validity of a deep learning algorithm to detect gait events from real-life walking bouts). Hidden Markov models, which segment a signal by assuming it moves through a fixed chain of hidden phases, split free-living Parkinsonian walking into strides at F1 92.1%. They also transfer from lab to home with little loss, but they degrade sharply in bouts under 15–30 strides (Hidden Markov Model based stride segmentation…). That last limitation constrains our design. Most domestic crutch use happens in short bouts, so our features must be computable over a few strides and pooled across bouts. They cannot depend on long steady walking.

Closest Prior Work

The Liar's Walk is the nearest neighbour. It classifies deceptive versus natural walking from video-derived 3D skeletons, combining 29 handcrafted posture and movement features with a long short-term memory network (a recurrent network that summarizes a sequence step by step). It reports 88.41% accuracy on its DeceptiveWalk dataset and 85.06% on unseen people. We differ in three ways. First, the deception is different. Those participants hid an intent (carrying a secret package) while walking normally, whereas ours fake a physical state. Second, the evidence is different. They used only kinematics from a fixed camera, while we use measured contact force and its timing against gait events. Third, the setting is different: one staged lobby session in their case, multi-day unsupervised wear in ours. We also test directly whether a flat learned sequence model of their kind matches our structural features. We treat that model as a serious rival, not a straw man.

Research Program

The hardware is an off-the-shelf forearm crutch with a single-axis load cell and an inertial measurement unit (IMU, a small accelerometer-plus-gyroscope package) in the shaft, plus one IMU on each shin. All sensors are clock-synchronized and log at 100 Hz or more to a phone. In the lab we validate the crutch load cell against a force plate. Foot contacts come from the shin IMUs using a temporal convolutional network, a network that detects events with stacked filters over widening time windows, in the style of Romijnders et al.. We retrain it on our own reference data if pretrained weights are not available. Crutch strike, crutch peak load and crutch lift come from the shaft signals. Each bout then becomes an ordered stream of symbolic events with lags between them. This stream is our "unlexicalized sentence."

From this stream we compute three families of features. Magnitude features are peak crutch load as a fraction of body weight and loading rate. Flat structure features are counts of event bigrams and trigrams (which event follows which) plus the mean and stride-to-stride variability of each inter-event lag. Hierarchical structure comes from a small probabilistic grammar induced over events. It nests contact events into strides and strides into bouts, including pauses, turns and repeated corrective crutch placements. It is the direct motor analogue of the parse-tree rules in Feng et al.. We do not assume the hierarchy is needed. It is one arm of an ablation. If bigrams and lag statistics do as well, the paper's claim becomes the simpler one.

Ground truth comes from a design that does not need anyone to secretly deceive. Thirty-six able-bodied adults complete three conditions with a crutch in counterbalanced order. In the genuine constraint condition, a locked knee orthosis plus an unloading heel wedge produce real partial offloading. In the simulation condition, participants have just experienced the constraint and are then told to reproduce that gait without the brace, which makes simulation as good as we can reasonably ask for. The third condition is natural crutch walking. Each condition is recorded in a lab block, with and without a spoken serial-subtraction dual task, and then in two days of home wear per condition, with participants paid to keep the assigned pattern whenever they use the crutch. An external genuine group of about 20 patients prescribed partial weight-bearing after lower-limb fracture tests whether a detector trained on brace-induced constraint carries over to real injury. We expect a gap here, because pain-driven gait is not the same as braced gait, and we will report it honestly.

Baselines

The naive reference point is chance (50% on balanced simulated-versus-genuine windows) together with a majority-class predictor on the natural class mix. Every model is evaluated leave-subject-out, so none of them can succeed by recognizing individuals. The chess stylometry result shows that individual style is easy to fingerprint (McIlroy-Young et al.), so a subject-dependent split would inflate every number.

The strongest existing method is the fused handcrafted-plus-recurrent design of The Liar's Walk. We re-implement it on our sensor streams: shin and crutch IMU velocity, acceleration and jerk features, plus a three-layer recurrent network over raw windows. Its reported figure is 88.41% accuracy (85.06% on unseen participants) on DeceptiveWalk, but that is a different task and different data, so it is context, not a target. On our task there is no reported result found on Lacuna for simulated-versus-genuine impairment detection. The competing models are therefore judged against each other on our data. We give this baseline the crutch force channel as extra raw input, so any advantage of ours comes from representation and not from sensor access. All models share the same windows, the same subject folds and the same hyperparameter budget of 50 random-search trials each.

A second strong competitor is a classical gait-variability model. It is a logistic regression on stride-time coefficient of variation, left–right asymmetry and crutch peak-load variability. This reflects the established view that stride-to-stride variability is the main free-living gait biomarker (Sidoroff et al.).

The key ablations isolate our ingredient. One is magnitude-only features against structure-only features, where structure means the order of events and the lags between them with every force value removed, mirroring Feng's 78.5% word-free result. The other compares flat bigram and lag features with the hierarchical grammar. Beating the baselines means that structure-only features match or exceed the recurrent baseline in leave-subject-out area under the ROC curve. It also means the full model beats it by a margin whose 95% subject-level bootstrap interval excludes zero. At home, where the ecological claim lives, the model must additionally reach a false-positive rate below 5% on genuine windows at a fixed operating threshold.

Why not just…?

Why not just train a deep sequence model on the raw signals? Such a model can in principle learn the coupling itself, and it is our strongest threat. The case against relying on it is not that it must fail. The concern is how it behaves with our data size, which is tens of participants, as in every study in this area. That is the same small-cohort limit that restricts Koontz et al. and Veličković et al.. At that scale, a flat recurrent model tends to latch onto magnitudes that simulators can copy, and onto individual style. The head-to-head comparison under identical folds settles it. If the recurrent model matches structural features at home and on the patient group, we will report that the explicit structure is unnecessary. Keeping structure-only features would then be justified only by interpretability.

Why not just use gait variability? A simulated limp may simply be noisier, and if so a two-feature variability model would be enough. We test this directly. The discriminating prediction is that genuine braced walking with a dual task is also more variable, while its lag structure stays intact. So variability should confuse "genuine but distracted" with "simulated," and structure should not. If the variability baseline performs as well across the dual-task blocks, the grammar framing is not needed and we will say so.

Why not just put pressure insoles under the affected foot? Insoles measure how much the foot is offloaded, and both genuine and simulated walkers offload. The shared quantity is exactly what the simulator is trying to reproduce. What insoles add is a second timing channel, so we include them in a sensor ablation (insoles alone, crutch alone, both) instead of dismissing them. If insole timing alone reproduces the result, the crutch load cell becomes optional. That would be a useful, cheaper finding.

Why not just measure dual-task cost on walking speed? Speed dropping under a cognitive load is a known sign of attention being taxed. It is a group-level effect that differs a great deal between individuals. We record speed and include dual-task speed cost as a single-feature baseline. We expect it to separate groups on average but to perform poorly per person, and the leave-subject-out numbers will show whether that expectation holds.

Decisive Checks

Three results would confirm the proposal, and each could fail. First, structure-only features must beat magnitude-only features in the lab. This is the motor analogue of the word-free grammar finding. Second, detection accuracy must rise from the lab block to home days and grow with hours since the day's first crutch use. If accuracy is flat or falls, the ecological claim is wrong. That would still leave a lab compliance-checking tool, but not the stronger claim. Third, a detector trained on braced participants must keep an acceptable false-positive rate on fracture patients. If it flags real patients as simulators, it must not be used for any purpose that affects individuals, and the paper will say this plainly. As a secondary check that the structural representation is not tied to our hardware, we rebuild hand–foot event order from the skeletons in DeceptiveWalk. We then ask how much of the 88.41% reported there a structure-only model recovers. That case involves hidden intent rather than faked impairment, so a low number there would not count against the main claim.