Wearable sensors find repeatable gait motifs in some children with cerebral palsy
Crouch gait in cerebral palsy is clinically familiar and mechanically unforgiving. The child walks with excessive knee flexion, and the pattern can vary enough from one stride to the next to complicate both assessment and device design. This study asked a practical question with direct relevance to assistive robotics: are there stable kinematic motifs in crouch gait that a powered exoskeleton could recognize and track reliably?
Investigators studied five children with cerebral palsy, aged 8 to 14 years and classified as Gross Motor Function Classification System levels I to III, while they walked at self-selected speed. Each child wore bilateral shank inertial measurement units, knee goniometers, and insole pressure sensors placed at the heel and forefoot. The signal stack was then filtered with a two-state Kalman filter, segmented into gait cycles using plantar-pressure events, and normalized to 101 points per cycle. That pipeline matters because any control strategy for a powered orthosis lives or dies on the quality of its input signal; if the gait pattern cannot be consistently detected, the device cannot respond with any confidence.
Knee motion proved far less repeatable than foot loading
The core result was not uniform across body segments. Repeatability tracked the signal measured rather than the functional severity of the child’s gait disorder. Mean ICC(3,1) reached 0.76 for forefoot loading, 0.61 for heel loading, and only 0.29 for knee kinematics. In plain terms, the pressure pattern under the foot was often stable enough to define a recognizable cycle, while the knee angle signal was much more erratic.
Five of the six failures were knee channels. Those same channels were the only ones that could not be distinguished from a permuted null, which suggests that the apparent knee pattern in several limbs may not have been robust enough to separate true structure from random variation. One child stood apart: in the least affected participant, knee signals met all criteria bilaterally. A single child cannot establish a general rule, but the contrast is clinically important because it points to heterogeneity within a diagnosis that is often treated as mechanically uniform.
Motif detection depended on a stricter signal test than simple symmetry
The study defined a kinematic motif as a limb’s ensemble waveform and then validated it using a multi-part framework: coefficient of variation, ICC(3,1), dynamic time warping, autocorrelation, and 1,000-iteration permutation testing. Symmetry was quantified with the Symmetry Index. This is a more demanding approach than simply asking whether two limbs look similar on a graph. It tests whether a waveform is repeatable enough across cycles to justify being called a motif rather than a noisy average.
Across the 30 limb-signal combinations, 12 met all criteria, 12 satisfied one or two, and 6 met none. That spread shows how uneven crouch gait can be when examined at the level of sensor-derived biomechanics. Forefoot and heel pressure measures were more likely to hold together across cycles than knee motion, a finding that carries obvious implications for exoskeleton control strategies that rely on detecting stance and transition phases. A device anchored to foot-contact signals may have a more dependable foundation than one that depends primarily on knee-angle regularity.
Why foot pressure may be a better control signal than knee kinematics
The likely reason is mechanical rather than mysterious. In crouch gait, knee flexion is shaped by muscle weakness, lever-arm dysfunction, balance compensation, and variable loading patterns across the limb. The foot still has to contact the ground, accept weight, and transition through stance, so plantar-pressure events may retain a more repeatable temporal signature even when joint angles fluctuate. Knee kinematics, by contrast, can shift with fatigue, compensatory trunk motion, or subtle changes in step-to-step strategy.
That biological and mechanical mismatch explains why the forefoot channel outperformed the knee channel in this small cohort. It also explains why powered exoskeletons are not simply waiting for a more sensitive sensor; they need the right signal from the right part of the gait cycle. For children with cerebral palsy, especially those whose crouch pattern looks irregular to the naked eye, the most dependable control input may be the one tied to ground contact rather than the one tied to joint posture.
A feasibility study, not a classifier trial
The study establishes feasibility in a highly heterogeneous cohort, but its boundaries are clear. Only five children were examined, and no gait-phase classifier was trained or evaluated. The work therefore stops short of showing that a device can autonomously identify phase transitions in real time, much less adapt its assistance across children with different motor profiles.
Even so, the findings are useful because they identify where future engineering effort should go. Knee signals may still have value, but they will likely need better filtering, individualized modeling, or integration with other sensors before they become dependable enough for powered orthotic control. The fact that the least affected child showed bilateral knee motif stability hints that severity and signal regularity may diverge within cerebral palsy, so any control algorithm will probably need to be personalized rather than built around a single gait template.
Funding came from Manipal Academy of Higher Education and the Science and Engineering Research Board in India as part of a broader project on an affordable therapeutic solution for children with crouch gait. The present study does not prove clinical benefit, yet it does define a usable engineering question: which wearable signals are stable enough to anchor control in a disorder marked by step-to-step variability? Here, plantar-pressure events looked far more reliable than knee kinematics, and that asymmetry will likely shape the next round of exoskeleton design.