Posture-Conditioned Frequency Analysis of Depth-Derived Finger Pose

A system that acquires 3D finger joint positions from consumer depth sensors and performs posture-conditioned spectral analysis to extract clinically meaningful motor biomarkers.

Method Summary

Six-stage processing pipeline designed to run on mobile hardware at 30 fps with total latency under 50 ms.

1

Depth Acquisition & Normalisation

From LiDAR, stereoscopic, or monocular sources into canonical 256×192 representation

2

Hand Pose Estimation

21-joint canonical skeleton via depth-based regression with kinematic constraints

3

Posture State Classification

Into clinically defined states: rest, postural hold, kinetic, intention

4

Pose-Anchored Residual Extraction

Per-finger local coordinate frame eliminates voluntary motion contamination

5

Frequency Feature Computation

Feature matrix indexed by posture state, joint, and spectral descriptor

6

Validity-Gated Aggregation

Assessment indicators with propagated confidence estimates

Key Innovation

The method addresses a fundamental limitation of existing tremor measurement: the failure to condition frequency-domain features on the postural context during which data is acquired.

Canonical Skeleton & Pose-Anchored Frames

21-Joint Hand Model

Wrist root (501) with five kinematic chains:

  • → Thumb: CMC → MCP → IP → Tip (502-505)
  • → Index: MCP → PIP → DIP → Tip (506-509)
  • → Middle: MCP → PIP → DIP → Tip (510-513)
  • → Ring: MCP → PIP → DIP → Tip (514-517)
  • → Little: MCP → PIP → DIP → Tip (518-521)

3D Position Recovery

Zj = DN(uj, vj) + Δdj
Xj = (uj − cx) · Zj / fx
Yj = (vj − cy) · Zj / fy

Local Coordinate Frame

For each finger k, a local frame Fk(t) is defined with:

Origin
MCP joint position PMCP(t)
z-axis (longitudinal)
Unit vector MCP → PIP
x-axis (lateral)
Perpendicular to z, normalized

Why this works: When the hand drifts voluntarily, both PMCP(t) and Rlocal(t) change, absorbing voluntary motion. The residual δlocal(t) captures only involuntary tremor without coordinated proximal joint changes.

Signal Processing

Short-time Fourier transform with Hann windowing and posture-conditioned feature extraction.

STFT Parameters

Window length2 seconds (60 samples)
Frequency resolution~0.47 Hz
Window advance0.25 sec (87.5% overlap)
Analysis band3-15 Hz

Spectral Features (per posture state)

f_peak(s)
Peak frequency
argmax_f PSD(f) for f ∈ [3, 15] Hz
P_tremor(s)
Tremor power
∫₃¹⁵ PSD(f) df
A_tremor(s)
Tremor amplitude
√(2 · P_tremor) — RMS displacement in mm
SC(s)
Spectral concentration
∫_{f_peak−1}^{f_peak+1} PSD(f) df / P_tremor
HR(s)
Harmonic ratio
PSD(2·f_peak) / PSD(f_peak)
γ(s)
Inter-finger coherence
|C_{jk}(f_peak)|² between fingers j and k

Kinetic Task Features

ADI
Amplitude decrement
Linear regression slope of peak-to-peak amplitude vs cycle number
FDI
Frequency decrement
Linear regression slope of inter-tap interval vs cycle number
HC
Hesitation count
Cycles where movement arrests (< 10% mean amplitude for > 100 ms)

Cross-Posture Composite Features

Rest-to-Postural Ratio (R/P)
R/P = A_tremor(Rest) / A_tremor(Postural)
R/P >> 1: Parkinson's pattern | R/P << 1: Essential tremor pattern

Quality Metrics & Confidence Propagation

LiDAR (dToF)

Base confidence α1.0
Depth noise σ_d±1-3 mm

Stereo (SGM)

Base confidence α0.7-0.9
Depth noise σ_d±3-10 mm

Monocular (ML)

Base confidence α0.6
Depth noise σ_d±10-30 mm

Validity Gating Criteria

C1: Depth Confidence
Mean per-pixel confidence in hand region > threshold
C2: Landmark Confidence
Fingertip localization confidence > threshold for all analyzed joints
C3: Motion Artifact
Low-frequency wrist displacement < 10 mm within analysis window

A window is validated if and only if all three criteria are satisfied throughout the duration.

Impact: In synthetic experiments, ungated estimates show 40-60% reduction in spectral concentration and up to 2 Hz peak frequency displacement. After gating, spectral concentration recovers to within 5% of artifact-free reference.

Planned Validation Studies

Test-Retest Reliability

Participants
N ≥ 30 healthy controls
Primary Endpoint
ICC for extracted features
Target
ICC > 0.80 for frequency, amplitude, concentration

Concurrent Validity

Participants
N ≥ 60 patients (PD, ET, mixed)
Primary Endpoint
Correlation with MDS-UPDRS Part III and Fahn-Tolosa-Marin scale
Target
Classification accuracy (PD vs ET vs control)

Cross-Modality Consistency

Participants
N ≥ 20 subjects on 3 devices
Primary Endpoint
Concordance correlation between modalities
Target
Validate cross-sensor normalization

Longitudinal Monitoring

Participants
N ≥ 15 PD patients over 12 weeks
Primary Endpoint
Correlation between progression slope and clinical ratings
Target
Detect medication adjustment effects within 1 week

Reference Datasets

  • →PhysioNet — Gait in Parkinson's Disease Database
  • →MJFF — Levodopa Response Study with wearable sensors
  • →Ninapro — EMG and kinematic data for hand movements
  • →FreiHAND / InterHand — Hand pose estimation benchmarks

Intellectual Property

Patent protection covering three independent claims across two statutory categories.

Core Method

Computer-Implemented Method

Posture-conditioned frequency analysis with depth acquisition, 3D joint pose estimation, posture state classification, validity gating, and feature computation from pose-anchored local coordinate frames

Guided Assessment Protocol

Independent Method Claim

Real-time compliance verification with posture matching, stability criteria, range-of-motion validation, color-coded feedback, and automated task repetition

System Architecture

System Claim

Complete pipeline combining pose-anchored frames, multi-criteria validity gating, posture-conditioned analysis, and confidence-scored outputs on mobile devices

Novelty Position

While prior art covers individual components (hand pose estimation, tremor detection, smartphone assessments), the novelty lies in the specific combination:

  • ✓Posture-conditioned frequency analysis using depth-derived finger pose
  • ✓Pose-anchored residual displacement in per-finger local coordinate frames
  • ✓Multi-criteria validity gating calibrated to depth sensor artifacts
  • ✓Cross-sensor normalization with confidence propagation

Freedom to Operate

Claims are framed as technical processes operating on sensor data with concrete technical effects (noise rejection, artifact exclusion, measurement reliability). Outputs are assessment indicators, not diagnoses, addressing EPC Article 53(c) exclusions.

Publications

Work in progress. Manuscripts in preparation for:

• IEEE Transactions on Biomedical Engineering
• npj Digital Medicine
• Movement Disorders / Neurology