Hypothesis

Passive acoustic emissions generated during active range-of-motion exercises, recorded via standard smartphone microphones and processed through convolutional neural networks operating on mel-frequency spectrogram representations, can detect subclinical articular cartilage degradation in rheumatoid arthritis (RA) joints 8–16 weeks before ultrasonographic evidence of erosion progression.

Background

Joint crepitus — audible and sub-audible acoustic signals generated by articular surfaces during movement — contains information about cartilage integrity, synovial fluid viscosity, and periarticular tissue mechanics. While clinicians have long used crepitus as a qualitative sign, systematic acoustic analysis remains underexplored. Recent advances in digital signal processing and deep learning applied to biomedical acoustics (joint sounds, lung auscultation, cardiac signals) suggest that smartphone-grade microphones (sampling at 44.1–48 kHz) capture sufficient spectral resolution to discriminate pathological from physiological joint sounds.

Cartilage degradation in RA produces characteristic changes in acoustic emission profiles: increased high-frequency components (2–8 kHz) from roughened articular surfaces, altered temporal patterns reflecting irregular motion, and loss of the smooth spectral envelope seen in healthy joints. These acoustic changes may precede structural damage visible on imaging.

Proposed Method

1. Acquisition: Standardized 30-second active flexion-extension recordings of MCPs, wrists, and knees using smartphone microphone in a quiet environment, with the device placed 2–5 cm from the joint
2. Preprocessing: Bandpass filter (200 Hz–12 kHz), ambient noise subtraction via adaptive spectral gating, segmentation into individual movement cycles
3. Feature extraction: Mel-frequency spectrograms (128 bands, 25ms windows, 10ms hop) as 2D input tensors
4. Classification: ResNet-18 backbone pre-trained on AudioSet, fine-tuned on joint acoustic spectrograms with binary output (progressor vs. non-progressor)
5. Validation: Prospective cohort (n≥200 RA patients), monthly acoustic recordings with quarterly musculoskeletal ultrasound (MSUS) as ground truth for erosion progression (OMERACT-EULAR scoring)

Testable Predictions

• Acoustic classifier AUC ≥0.80 for predicting MSUS erosion progression at 16 weeks
• Sensitivity ≥75% at specificity ≥70% for subclinical progression detection
• Acoustic spectral entropy increases ≥0.3 SD in progressor joints vs. non-progressors at 8 weeks pre-erosion
• Model performance generalizes across smartphone models (iPhone/Android) with AUC drop <0.05
• Combined acoustic + CRP + anti-CCP model outperforms imaging-only monitoring (net reclassification improvement >0.15)

Limitations

• Ambient noise contamination in non-clinical settings may reduce signal quality
• Body habitus (subcutaneous tissue thickness) affects acoustic transmission
• Crepitus is not specific to cartilage damage — synovitis, tendinopathy, and gas cavitation produce overlapping signals
• Smartphone microphone variability across manufacturers introduces hardware confounding
• Training data requirements may be substantial given acoustic heterogeneity across joint sites
• Longitudinal validation needed to establish temporal stability of acoustic biomarkers

Clinical Significance

If validated, smartphone-based joint acoustic monitoring would provide a zero-cost, patient-administered, infinitely scalable screening tool for RA structural progression. This could enable remote monitoring between clinic visits, reduce unnecessary imaging, democratize access to early erosion detection in resource-limited settings, and support treat-to-target strategies by providing continuous rather than episodic structural assessments. Integration with existing digital health platforms would require minimal infrastructure.

LES AI • DeSci Rheumatology