📄 health informatics

Supervised Contrastive Learning-based Digital Biomarker Discovery for Wearable IMU Gait Signals

This study introduces the Embedding-Distance Gait Biomarker (EDGB), a supervised contrastive learning framework that utilizes a compact convolutional neural network to extract robust 32-dimensional latent representations from raw wearable IMU signals, achieving high accuracy in distinguishing between healthy, neurological, and orthopedic gait patterns while demonstrating strong reliability and significant group differentiation.

Original authors: Mohtavipour, S. M.

Published 2026-07-04

📖 4 min read☕ Coffee break read

CC BY 4.0

Original authors: Mohtavipour, S. M.

Original paper licensed under CC BY 4.0 (https://creativecommons.org/licenses/by/4.0/). ⚕️ This is an AI-generated explanation of a preprint that has not been peer-reviewed. It is not medical advice. Do not make health decisions based on this content. Read full disclaimer

Imagine you are trying to tell the difference between three types of walkers: a healthy person, someone with a neurological condition (like Parkinson's), and someone with an orthopedic issue (like a knee injury).

Traditionally, doctors might watch you walk and guess, or they might use a stopwatch to count your steps and measure how long it takes you to turn around. The problem is that these "stopwatch" measurements only look at one or two things at a time, like a single note from a piano, rather than the whole song. They often miss the subtle, complex ways our bodies move.

This paper introduces a new "digital biomarker" called EDGB (Embedding-Distance Gait Biomarker). Think of this not as a stopwatch, but as a high-tech GPS for your walking style.

Here is how it works, broken down into simple steps:

1. The Sensors: The "Ears" on Your Body

The researchers used small wearable sensors (IMUs) placed on four spots: the head, the lower back, and both feet. These sensors act like super-sensitive ears, listening to how your body moves. They don't just listen to where you are; they listen to:

Speed: How fast you are moving (acceleration).
Spin: How you are turning (angular velocity).
Changes: How quickly you speed up, slow down, or twist (called "jerk" and "angular acceleration").

2. The Brain: The "Translator"

The raw data from these sensors is messy and hard to read. The researchers built a small, smart computer program (a neural network) to act as a translator.

The Training: They showed this program thousands of walking examples from healthy people, neurological patients, and orthopedic patients.
The Lesson: Using a technique called "Supervised Contrastive Learning," the program learned a very specific rule: "Group all the healthy walkers together in one corner of the room, all the neurological walkers in another corner, and all the orthopedic walkers in a third corner."
The Result: The program learned to compress a complex 19-second walk into a tiny, 32-number "fingerprint" (an embedding) that perfectly captures the essence of that person's gait.

3. The Biomarker: The "Distance Score"

Once the program is trained, it creates a "map" with three central points (prototypes): one for the average healthy walker, one for the average neurological walker, and one for the average orthopedic walker.

When a new person walks, the system creates their fingerprint and asks: "How far is this person from the healthy center? How far from the neurological center? How far from the orthopedic center?"

The EDGB is simply a single number calculated from these distances.

If the number is low, the person walks like a healthy person.
If the number is high, they walk like someone with an orthopedic issue.
If it's in the middle or negative, they might walk like someone with a neurological issue.

Why is this better than the old way?

The researchers tested this new "GPS" against the old "stopwatch" methods (like measuring stride time or turn speed).

The Stopwatch: It was okay at telling healthy people apart from sick people, but it got very confused when trying to tell the difference between a neurological patient and an orthopedic patient. It was like trying to tell a violin from a cello by only listening to the volume.
The GPS (EDGB): It was incredibly sharp. It could tell the difference between neurological and orthopedic walkers with 99.5% accuracy. It was like listening to the entire orchestra and instantly knowing which instrument was playing.

The "Secret Sauce"

The paper found that the most important information came from the feet and from measuring how fast the movement changed (jerk), not just the movement itself. It's like realizing that to understand a dancer, you don't just watch where their feet land; you watch how quickly they lift and drop them.

Does it work consistently?

Yes. The researchers checked if the score changed every time the same person walked. They found that the score was very stable (82% consistent). This means the number reflects the person's actual walking style, not just random wobbles or bad luck on a specific day.

Summary

This paper didn't invent a new drug or a new surgery. Instead, it invented a smarter way to measure walking. By using a smart computer to listen to the full "song" of a person's walk (including speed, spin, and sudden changes) rather than just counting steps, they created a single number that can accurately and objectively tell the difference between healthy walking, neurological walking, and orthopedic walking.

Technical Summary: Supervised Contrastive Learning-based Digital Biomarker Discovery for Wearable IMU Gait Signals

Problem Statement
Wearable inertial measurement units (IMUs) offer a practical method for gait assessment, yet traditional approaches relying on handcrafted features (e.g., stride time, cadence) often fail to capture the full multidimensional nature of pathological gait patterns. While deep learning models can learn complex patterns from raw signals, they frequently function as "black boxes" with limited interpretability and may suffer from information leakage when evaluation protocols use trial-wise splits rather than subject-wise splits. Furthermore, existing methods often lack a unified, compact digital biomarker that can objectively distinguish between healthy, neurological, and orthopedic gait patterns while maintaining clinical relevance.

Methodology
The study proposes a digital biomarker termed the Embedding-Distance Gait Biomarker (EDGB), derived from supervised contrastive representation learning. The methodology involves the following components:

Dataset: The research utilizes the publicly available Voisard clinical gait dataset, comprising 1,356 gait trials from 260 participants (73 healthy, 143 neurological, 44 orthopedic). Data was collected using four synchronized IMUs (head, lower back, left foot, right foot) recording 3-axis acceleration and angular velocity at 100 Hz.
Input Preprocessing: The network accepts 16 input channels per trial. These include the magnitude of acceleration ( $\|a\|$ ), angular velocity ( $\|\omega\|$ ), and their first temporal derivatives (jerk $\|j\|$ and angular acceleration $\|\alpha\|$ ) from all four sensors. All signals are resampled to a fixed length of 1,900 samples (~19 seconds).
Deep Learning Architecture: A compact 1D Convolutional Neural Network (CNN) is employed as an encoder. It consists of three convolutional layers (with 4, 8, and 16 channels) followed by batch normalization, ReLU activations, and max-pooling. The output is processed via adaptive average and max pooling to generate a 256-dimensional feature vector. This is mapped to a 32-dimensional latent embedding space ( $e$ ) via a fully connected embedding network. A projection head ( $z$ ) is used during training but discarded during inference.
Training Objective: The model is trained using Supervised Contrastive Learning (SupCL). Unlike standard classification, SupCL constructs positive pairs from trials of the same clinical category and negative pairs from different categories within a batch. The loss function minimizes the distance between embeddings of the same class while maximizing the distance between different classes in the latent space.
Biomarker Construction: After training, class-specific prototype vectors ( $\mu_H, \mu_N, \mu_O$ ) are computed as the normalized centroids of the training embeddings for Healthy, Neurological, and Orthopedic groups. For a new trial with embedding $e$ , the EDGB is calculated as a scalar value based on the Euclidean distances to these prototypes: $DBM = d_H + d_N - d_O$ . This formulation is designed to maximize separation, where lower values indicate proximity to the neurological cluster and higher values indicate proximity to the orthopedic cluster.
Evaluation Protocol: A strict subject-level split (80% training, 20% testing) is used to prevent data leakage from repeated trials of the same participant.

Key Contributions

Novel Biomarker Definition: The introduction of EDGB, a scalar digital biomarker derived from the distance of a gait embedding to learned class prototypes, providing a compact representation of complex gait dynamics.
Supervised Contrastive Framework: The application of SupCL to wearable IMU data to learn a discriminative latent space that clusters clinical categories effectively without relying on manual feature engineering.
Integration of Higher-Order Dynamics: The explicit inclusion of jerk and angular acceleration derivatives as input channels, demonstrating their critical role in capturing movement dynamics.
Rigorous Validation: Evaluation on unseen subjects with a focus on repeatability (ICC) and statistical significance (ANOVA), ensuring the biomarker reflects subject-specific characteristics rather than trial-to-trial noise.

Results

Discriminative Performance: On the held-out test set, the EDGB achieved an overall accuracy of 84% in classifying subjects into the three clinical groups.
- Healthy vs. Neurological: AUC = 90.59%
- Healthy vs. Orthopedic: AUC = 88.47%
- Neurological vs. Orthopedic: AUC = 99.50% (near-perfect separation)
- The macro-average AUC for EDGB was 92.85%, significantly outperforming the best conventional handcrafted feature (U-turn time), which achieved a macro-AUC of 76.48%.
Statistical Significance: ANOVA analysis revealed that clinical category explained 71% of the variance in the EDGB ( $\eta^2 = 0.7104$ ), a substantially higher effect size than any conventional gait measure (the next highest was U-turn time at $\eta^2 = 0.1628$ ).
Reliability: The biomarker demonstrated high repeatability across trials with an Intraclass Correlation Coefficient (ICC(2,1)) of 0.82, indicating that 82% of the variance is attributable to between-subject differences.
Interpretability: Correlation analysis showed weak associations with conventional timing features (e.g., stride time, cadence) but moderate-to-strong correlations with jerk-related signal features (e.g., maximum jerk, jerk variability), particularly from foot-mounted sensors. This suggests the biomarker captures dynamic movement characteristics beyond simple temporal metrics.
Ablation Studies:
- Sensor Importance: Foot-mounted sensors (Right Foot and Left Foot) provided the most discriminative information.
- Input Channels: Including derivative signals (jerk and angular acceleration) improved the macro-AUC from 89.64% (raw signals only) to 92.85%.

Significance and Claims
The paper claims that the proposed EDGB framework successfully transforms multi-sensor IMU recordings into a compact, informative, and objective digital biomarker. The authors assert that the learned representation captures subtle, multidimensional gait characteristics specific to neurological and orthopedic disorders that are not adequately represented by traditional handcrafted descriptors. The study highlights that the biomarker is particularly effective at distinguishing between different pathological gait patterns (neurological vs. orthopedic), a task where conventional measures often struggle. The authors conclude that supervised contrastive learning offers an effective framework for developing clinically useful biomarkers for movement disorder assessment, provided that subject-level evaluation protocols are strictly adhered to. The study acknowledges limitations regarding the broad categorization of diseases and suggests future work should focus on finer-grained disease subtypes and longitudinal monitoring of disease progression.