Technical Summary: Using AI‑Assisted Centre of Mass Acceleration Analysis for Sacro Iliac Joint Dysfunction

Problem Statement

Clinical screening for Sacroiliac Joint (SIJ) dysfunction in runners currently relies heavily on brief visual gait observations and clinician-led checklists. While practical, these methods are constrained by observational subjectivity, inter-rater variability, and limited sensitivity to subtle biomechanical features, particularly within short gait samples. Traditional biomechanical models often simplify the pelvis as a rigid segment or rely on low-dimensional analyses (e.g., discrete joint angles), which may fail to capture the complex, multivariate coordination structures and nonlinear load transfer patterns characteristic of SIJ dysfunction. Furthermore, while wearable inertial measurement units (IMUs) provide rich data, the integration of Generative AI (GenAI) to interpret this data for SIJ-specific clinical decision-making remains underexplored.

Methodology

This study employed a prospective observational design involving a single experienced recreational male runner (39 years, 181 cm, 72 kg) with persistent unilateral (left-sided) SIJ dysfunction, diagnosed via clinical history and a cluster of pain provocation tests (including a positive FABER test).

Data Acquisition:

• Sensor: A Human Activity Monitor (HAM) with a triaxial accelerometer (100 Hz sampling rate) was affixed to the participant's posterior pelvis (approximating the S2 vertebra) to capture whole-body Centre of Mass (CoM) acceleration.
• Protocol: The participant underwent a graded return-to-running program over several weeks. Data was collected during 10 x 60-second running trials on an athletics track.
• Processing: Raw acceleration data from the first and last 10 seconds of each trial were extracted. Time-domain features were manually processed offline, including Root Mean Square (RMS) acceleration, peak absolute acceleration, step-to-step variability (Coefficient of Variation), and symmetry indices (Left vs. Right).

AI Integration:

• Tool: Microsoft 365 Copilot (GenAI) was utilized as a Clinical Decision Support System (CDSS).
• Process: Raw CoM acceleration data and specific biomechanical prompts (focusing on mediolateral control, load transfer, and symmetry) were fed to Copilot. The AI was tasked with:
  1. Identifying multivariate biomechanical features associated with SIJ dysfunction.
  2. Predicting movement signatures related to asymmetry and load management.
  3. Synthesizing findings to support "return-to-running" readiness.
• Comparison: AI-derived metrics and interpretations were compared against a clinical observational checklist completed by a physiotherapist and biomechanist, focusing on visible asymmetries (pelvic drop, trunk shift) and gait irregularities.

Key Contributions

• AI-Augmented Clinical Reasoning: The study demonstrates a framework where GenAI (Copilot) acts as an adjunct to, rather than a replacement for, clinical judgment. It processes raw, short-duration acceleration data to extract objective metrics that complement subjective visual assessments.
• Multivariate Feature Extraction: The research highlights the ability of GenAI to identify subtle, non-linear acceleration characteristics (e.g., mediolateral variability, oscillatory patterns) that are difficult to detect via checklist-based visual observation alone.
• Bridging Data and Diagnosis: The paper establishes a method for mapping raw accelerometer features (RMS, peak ratios, harmonic ratios) to specific clinical constructs of SIJ dysfunction, such as impaired load transfer and pelvic instability.

Results

• AI Classification: Copilot successfully differentiated gait states corresponding to early rehabilitation, transitional recovery, and return-to-running readiness. It identified a "left-sided control and loading pattern" characterized by:
  • Mediolateral (ML) Axis: Pronounced asymmetry (10.2% RMS asymmetry) and elevated variability (5.5% CV), indicating reduced pelvic stability and altered trunk lean on the affected side.
  • Vertical (VT) Axis: Reduced consistency and increased impact attenuation variability, with the affected side showing smaller vertical peaks (indicating unloading) compared to the contralateral side.
  • Anteroposterior (AP) Axis: Asymmetric braking/propulsion forces and moderate stride-to-stride variability.
• Clinical Correlation: The AI-derived metrics aligned with clinical observations of pelvic drop, lateral trunk motion, and diminished single-leg balance.
  • Statistical Agreement: There was moderate agreement between clinical concern and Copilot risk classification (Weighted Cohen's κ = 0.48).
  • Correlation: Clinically observed step inconsistency showed a strong positive association with AI-derived gait variability, while pelvic impairment showed a moderate inverse association with AI-derived symmetry indices.
• Fatigue Simulation: The AI analysis revealed that as the running session progressed (simulating fatigue), the unloading of the left limb increased, and contralateral overload worsened, a trend consistent with clinical observations of pain provocation.

Significance and Claims

The paper claims that the integration of GenAI with wearable CoM acceleration data provides an objective, reproducible complement to traditional clinician-led gait observation for SIJ dysfunction screening.

• Augmentation, Not Replacement: The authors explicitly state that the study does not aim to replace clinicians or make definitive diagnoses. Instead, GenAI serves as a "biomechanical second lens" that strengthens clinical reasoning by providing continuous, quantitative feedback on movement quality and symmetry.
• Scalability: The approach offers a potential pathway toward remote, scalable, and accessible gait assessments in clinical and home environments, supporting more consistent decision-making in return-to-running protocols.
• Modest Scope: The authors acknowledge the limitations of the single-case design, noting that findings are not generalizable without further validation. They emphasize that AI outputs are most effective when integrated with expert judgment and that the study serves as a proof-of-concept for a clinical framework rather than a validated predictive model for broad populations.