⚡ electrical engineering

Early Prediction of Imminent Intrinsically Generated Trips During Gait in People Post-Stroke

This study developed and validated an early, targeted predictor using a two-feature linear model that successfully identifies 88% of imminent intrinsically generated trips in people post-stroke approximately 191 ms before trip onset, enabling timely and necessary real-time interventions to prevent falls.

Original authors: Austin Louis Mituniewicz, He (Helen) Huang, John M. Baratta, Michael D. Lewek

Published 2026-08-18

📖 5 min read🧠 Deep dive

CC BY 4.0

Original authors: Austin Louis Mituniewicz, He (Helen) Huang, John M. Baratta, Michael D. Lewek

Original paper licensed under CC BY 4.0 (https://creativecommons.org/licenses/by/4.0/). ✨ This is an AI-generated explanation of the paper below. It is not written or endorsed by the authors. For technical accuracy, refer to the original paper. Read full disclaimer

For millions of people living with the aftermath of a stroke, the simple act of walking carries a hidden danger. While many assume that falling is a result of losing balance or tripping over an object, a significant number of these accidents happen because of a glitch within the body's own movement. As a person walks, their leg swings forward in a smooth arc. Sometimes, without any external obstacle, the foot fails to lift high enough and scrapes against the floor. This internal stumble, known as an intrinsically generated trip, can cause the leg to catch, throwing the person off balance and often leading to a fall. For stroke survivors, such falls are particularly devastating, frequently resulting in broken bones, loss of independence, and a deep-seated fear of walking again. While physical therapy helps many, it does not always stop these trips, and the body's natural reaction to a stumble is often too slow to prevent a fall. This leaves a critical gap: how can technology intervene quickly enough to stop a trip before it happens, without constantly interfering with normal walking?

A team of researchers at the University of North Carolina at Chapel Hill and North Carolina State University set out to build a system that could predict these internal trips before they occur. Their goal was to create a method that acts like a smart safety net, stepping in only when necessary. They recruited eleven individuals who had experienced a stroke at least six months prior and had reported at least one trip or near-fall in the previous six months. To study these events in a controlled environment, the participants walked on a treadmill while researchers gently challenged their focus. The participants were asked to look away from their feet, let go of handrails, or perform mental tasks like naming foods, simulating the distractions that often lead to real-world accidents. As they walked, the researchers used motion-capture cameras to track the precise movements of their legs and recorded the forces their feet made against the treadmill belt.

The challenge for the researchers was that these trips are rare. In the thousands of steps the participants took, only a small fraction resulted in a trip or a significant scrape against the floor. To train a computer to recognize these rare events, the team had to teach it to spot the subtle warning signs that appear just before a foot catches. They focused on the moments just before the foot leaves the ground, a phase of walking called late stance. By analyzing the angle of the lower leg and the speed of the swing, they looked for patterns that differed from a normal, smooth step. They grouped the data into two categories: normal steps and "abnormal" steps, which included both full trips and those heavy scrapes where the foot hit the ground but the person managed to recover.

The researchers tested two different ways to help the computer learn from this unbalanced data, where normal steps vastly outnumbered the dangerous ones. One method involved removing some of the normal steps to make the groups more equal, while the other method simply told the computer to treat a mistake on a dangerous step as a much bigger error than a mistake on a normal step. The second approach proved more effective. The computer learned that the most reliable warning sign was the angle of the lower leg just before it lifted off the ground. When a trip was about to happen, the lower leg tended to be positioned more horizontally, or flatter, than usual. By watching for this specific posture, the system could identify a potential trip with high accuracy.

The most significant finding was how early the system could make this prediction. The computer could flag a potential trip an average of 191 milliseconds before the trip actually began. This is a crucial window of time. It is significantly earlier than the moment the foot leaves the ground, which is when many current devices, such as electrical stimulators, typically activate. Because muscles take time to react to an electrical signal, waiting until the foot is already off the ground is often too late to prevent a fall. By predicting the trip nearly 200 milliseconds in advance, the system provides enough time for a device to stimulate the leg muscles and lift the foot higher, clearing the ground before the catch occurs. The system was also tested on participants whose data was not used to build the model, and it performed well, correctly identifying trips in five out of eight participants and predicting nearly 90 percent of all trips overall.

While the study was conducted in a laboratory setting with a limited number of participants, the results suggest a promising path forward. The researchers did not build a wearable device that people could use today, but they demonstrated that the data needed to predict a trip exists and can be found using simple sensors that track leg movement. The work highlights that the key to preventing these falls lies not in constant assistance, which can be tiring and unnecessary, but in precise, timely intervention. By understanding the specific kinematic signatures of a trip before it happens, engineers can design future assistive technologies that act only when the body needs them, offering a safer, more independent way to walk for those recovering from a stroke.

Technical Summary: Early Prediction of Imminent Intrinsically Generated Trips During Gait in People Post-Stroke

Problem Statement
Falls in people post-stroke frequently result from intrinsically generated trips (IGTs)—unexpected contacts between the foot and ground during the swing phase without external obstacles. While physical training is a common intervention, it often fails to reduce fall rates in this population due to delayed neuromuscular responses and voluntary activation deficits. Existing engineering solutions, such as ankle-foot orthoses (AFOs) and functional electrical stimulation (FES) foot-drop stimulators, operate on a "always-on" basis. This approach can be restrictive, cause muscle fatigue, and, critically, may engage too late to prevent a trip due to the electromechanical delay (EMD) between muscle excitation and force generation. Current prediction methods often rely on person-specific models requiring prior collection of trip data or depend on discrete gait events (like foot-off) that limit the time available for intervention. The challenge is to develop a generalized, early prediction system that operates in real-time using limited kinematic sensors, providing sufficient lead time for an intervention to act before the trip occurs.

Methodology
The study involved 11 individuals with chronic stroke (hemiparesis) walking on a split-belt instrumented treadmill. To induce IGTs, researchers employed non-environmental distractors (e.g., cognitive tasks, head turns) and varied walking speeds. Data were collected using a 10-camera motion capture system and instrumented treadmill at 120 Hz and ≥960 Hz, respectively.

Data Labeling and Class Definition: Researchers visually identified 58 IGTs from 8 of the 11 participants. To address the severe class imbalance (112:1 ratio of normal to IGT strides), the "abnormal" class was expanded to include both IGTs and "large scuffs" (strides with peak vertical ground reaction force [vGRF] ≥ 20 N during swing but no visible kinematic deviation). This reduced the imbalance to 22:1.
Feature Extraction: Features were derived solely from paretic-side thigh and shank sagittal plane kinematics (angular positions, velocities, accelerations) and a linear phase estimator ( $\Phi$ ). Data were segmented into nine overlapping windows defined by the continuous gait phase estimator during late stance and pre-swing ( $\Phi \in (25\%, 50\%]$ ).
Model Development: A Sequential Forward Selection (SFS) algorithm was used to select features for a linear kernel Support Vector Machine (SVM). Two methods for handling class imbalance were tested: NearMiss3 (a data-level undersampling method) and a cost-sensitive "balanced-cost" approach.
Early Prediction Strategy: To enable prediction before the end of the analysis window, a counter-based approach was implemented. The model counts the number of samples labeled as "abnormal" within the final window ( $\Phi \in (45\%, 50\%]$ ). A threshold was established to trigger a prediction once a specific number of abnormal samples were detected, allowing for intervention before the window concludes.
Validation: The system was evaluated using Leave-One-Participant-Out (LOPO) cross-validation. Performance was compared against a recreation of Zhang et al.'s person-specific outlier detection method using generalized features.

Key Results

Prediction Accuracy: The final model, utilizing the balanced-cost method and features from the $\Phi \in (45\%, 50\%]$ window (specifically shank angle and instantaneous cadence), achieved an average sensitivity of 88% (SD 0.17) and a specificity of 68% (SD 0.17) across the eight participants with trip data. This corresponds to a false alarm rate of 32%.
Timing: The system predicted IGTs significantly earlier than previous methods. The average prediction occurred 191 ± 77 ms prior to IGT onset. In contrast, the comparison method (Zhang et al.) made predictions at foot-off, which occurred only 86 ± 39 ms prior to IGT onset. The difference was statistically significant ( $p=0.008$ ).
Feature Significance: The shank angle in the late stance phase ( $\Phi \in (45\%, 50\%]$ ) was the strongest single predictor. Non-normalized data analysis revealed that the shank was significantly more horizontally oriented prior to an IGT ( $-31.8^\circ$ ) compared to normal strides ( $-27.8^\circ$ ).
Generalization: The model successfully predicted 100% of IGTs for five of the eight participants. In a preliminary validation with three participants excluded from training (who had no IGTs but varying levels of "large scuffs" or normal gait), the model maintained high accuracy, correctly predicting 96% of large scuffs for one participant and 100% of strides for another.
Comparison to State-of-the-Art: While the Area Under the Curve (AUC) of the new method (0.81 ± 0.15) was not statistically different from the recreation of Zhang et al.'s method (0.77 ± 0.16), the new method offered a significantly longer prediction horizon.

Significance and Claims
The authors claim this work represents a significant step toward the targeted prevention of trip-related falls in people post-stroke. The primary contribution is a generalized, early prediction algorithm that does not require prior collection of trip data from the specific user to function effectively. By predicting trips approximately 190 ms before onset (roughly 100 ms before foot-off), the system provides a time window that potentially exceeds the electromechanical delay of muscle stimulation. This suggests the method could be integrated with assistive technologies, such as FES, to intervene selectively only when a trip is imminent, thereby avoiding the drawbacks of continuous stimulation (fatigue, skin irritation) and restrictive bracing. The study emphasizes that while the offline results are encouraging, the method is designed with future real-time implementation in mind, utilizing data sources (thigh and shank kinematics) that can be acquired via inertial measurement units (IMUs).

Technical Summary: Early Prediction of Imminent Intrinsically Generated Trips During Gait in People Post-Stroke

More like this