Can AI Models Outpace Traditional Techniques in Gait Analysis?

The intersection of machine learning and gait analysis offers a captivating arena for exploration, yet it often stumbles over the barrier of interpretability. Enter large language models (LLMs), which some believe might bring a new level of explanation and confidence-aware outputs to kinematic analysis.

Evaluating LLMs in Gait Analysis

In a recent study, researchers sought to determine whether general-purpose LLMs could classify continuous gait kinematics when these are represented as textual numeric sequences. The study compared these language models, including GPT-5 and its variants, against traditional machine learning approaches like the supervised K-Nearest Neighbor (KNN) classifier and One-Class SVM (OCSVM), using a dataset from 20 participants with seven distinct gait patterns.

The results? The KNN classifier triumphed with a multiclass Matthews Correlation Coefficient (MCC) of 0.88, leaving the best-performing LLM, GPT-5, trailing with a multiclass MCC of 0.70. While the LLMs didn't quite match up, they did outperform the class-independent OCSVM, which had a binary MCC of 0.60.

LLMs: An Exploratory Tool or a Diagnostic Hope?

What does this mean for the future of clinical gait analysis? Color me skeptical, but the claim that LLMs could replace traditional classifiers doesn't survive scrutiny. The LLMs' performance heavily leaned on explicit reference information. In scenarios where predictions were filtered for high confidence, the multiclass MCC for LLMs climbed to 0.83. However, this reliance on reference grounding and confidence filtering raises questions about their reliability in clinical settings.

the computationally efficient o4-mini model performed comparably to its larger counterparts, but efficiency alone doesn't replace accuracy. The takeaway here's that while LLMs can provide intriguing insights, they remain exploratory tools. They're not ready to supplant supervised classifiers for precise gait classification without human oversight.

The Path Ahead

So, should we dismiss LLMs as mere curiosities in the clinical context? Not quite. Their ability to offer insights and aid exploration could still revolutionize how we approach gait analysis. However, their current iteration demands cautious, human-guided interpretation. The real question is whether future iterations of LLMs will bridge this gap between exploration and diagnostic capability.

As we push the boundaries of what AI can achieve, let's apply some rigor here. Until LLMs can consistently outperform traditional methods without significant human intervention, their role in clinical practice will remain supplementary at best. The challenge is clear: can LLMs evolve to become reliable diagnostic tools, or will they remain on the periphery, always a step behind?