Convolutional neural network for early detection of lameness and irregularity in horses using an IMU sensor

TL;DR

This work tackles the challenge of early, objective lameness detection in horses by employing a single IMU sensor and a 1D CNN to perform stride-level classification, with session-level decisions formed from an anomaly score. The method uses stride segmentation and a small set of multivariate time-series features collected from 42 horses, focusing on the trot gait for reliable discrimination. Key findings show that trot offers the strongest signal (75.5% stride-level accuracy) and that the system achieves 90% session-level accuracy with no false negatives, indicating strong potential for field deployment. The approach significantly reduces hardware complexity and cost while enabling timely, data-driven welfare and performance decisions in veterinary and equestrian practice.

Abstract

Lameness and gait irregularities are significant concerns in equine health management, affecting performance, welfare, and economic value. Traditional observational methods rely on subjective expert assessments, which can lead to inconsistencies in detecting subtle or early-stage lameness. While AI-based approaches have emerged, many require multiple sensors, force plates, or video systems, making them costly and impractical for field deployment. In this applied research study, we present a stride-level classification system that utilizes a single inertial measurement unit (IMU) and a one-dimensional convolutional neural network (1D CNN) to objectively differentiate between sound and lame horses, with a primary focus on the trot gait. The proposed system was tested under real-world conditions, achieving a 90% session-level accuracy with no false positives, demonstrating its robustness for practical applications. By employing a single, non-intrusive, and readily available sensor, our approach significantly reduces the complexity and cost of hardware requirements while maintaining high classification performance. These results highlight the potential of our CNN-based method as a field-tested, scalable solution for automated lameness detection. By enabling early diagnosis, this system offers a valuable tool for preventing minor gait irregularities from developing into severe conditions, ultimately contributing to improved equine welfare and performance in veterinary and equestrian practice.

Figures (9)

• Figure 1: Picture of the Alogo Move Pro sensor.
• Figure 2: Two drawings show the positioning and mounting of the sensor
• Figure 3: Trot stride length distribution of our dataset
• Figure 4: Overview of the 1D CNN model architecture for lameness detection.
• Figure 5: Comparison of test accuracy on strides for different gaits