PD Monitoring Glove

Sensing-to-Decision: A Low-Cost, Privacy-Centric Edge Framework for Objective Parkinsonian Tremor and Bradykinesia Quantification

Authors: An Nguyen, Madhu Babu Cherukuri, Vatsalya Rohitbhai Dabhi, Sarita Singh, Hongpeng Fu Lab: Intelligent Automation (IoT) Research Group

Overview

An IoT-enabled wearable glove system for continuous, objective monitoring of Parkinson's Disease motor symptoms. The system targets two primary biomarkers from the MDS-UPDRS Part III motor examination: resting tremor (4–6 Hz involuntary oscillation) and bradykinesia (progressive slowness and decremental amplitude during repetitive movement).

All sensor data is processed locally on a Raspberry Pi 5 through a DSP pipeline and Transformer-based ML model, with only processed clinical scores published via MQTT to the cloud. Raw biometric data never leaves the device (Privacy-by-Design).

Intended Use and Clinical Scope

• This prototype is intended for research-grade monitoring and decision support, not standalone diagnosis.
• MDS-UPDRS-aligned tremor tasks are used for structured data capture, but final clinical interpretation should remain clinician-supervised.
• For the standardized 3-task assessment protocol and separate therapeutic hand-exercise guidance, see docs/assessment-and-therapy-protocol.md.

Team Cloud Handoff (CSV -> AWS)

When sharing data with backend teammates, credentials and config are not enough on their own:

• Required: message schema (IoT payload fields/types/units) and storage schema (DynamoDB key/attribute mapping, optional S3 layout).
• Certificates/keys: enable secure connection only; they do not define payload structure.
• Security: never commit private keys/cert bundles to this repository.

System Architecture

[Glove Sensors]                   [Patient Phone Video]
5× MPU6050 IMU                    Encrypted session recording
5× Flex Sensor (thumb bench validated)  Post-session MediaPipe validation
        ↓                                  ↓
    ~89Hz measured (4ch), 100Hz target   Compliance flags (later stage)
        ↓
┌──────────────────────────────────────────┐
│   DSP: Butterworth Band-Pass (3-15 Hz)  │
│   + Fast Fourier Transform (FFT)        │
└──────────────────────────────────────────┘
        ↓
┌──────────────────────────────────────────┐
│   ML Inference (TFLite INT8)            │
│   Transformer Encoder → MDS-UPDRS 0-4   │
└──────────────────────────────────────────┘
        ↓
   MQTT JSON Payload → Cloud Web App

Hardware

Component	Qty	Purpose	Status
Raspberry Pi 5 (8GB)	1	Edge computing gateway	✅ Acquired
MPU6050 IMU (GY-521)	5	6-axis accel/gyro for tremor detection (one per finger)	✅ Acquired & Soldered
TCA9548A I2C Multiplexer	1	Routes I2C bus for 5 IMUs sharing address 0x68	✅ Acquired & Soldered
SparkFun Flex Sensor 2.2" (SEN-10264)	5	Resistive bend sensors for bradykinesia measurement	🧪 Bench validated (thumb only, off-platform on Arduino Nano — see docs/flex-bench-characterization.md); Pi 5 + MCP3008 integration pending
MCP3008 ADC	1	10-bit SPI ADC for flex sensor voltage dividers	✅ Acquired
4.7kΩ Pull-up Resistors	2	I2C SDA/SCL pull-ups	✅ Acquired
10kΩ Resistors	5	Flex sensor voltage divider pull-downs	✅ Acquired

Wiring & Sensor Placement

I2C Subsystem (Tremor Detection): Pi GPIO2 (SDA) / GPIO3 (SCL) → TCA9548A (0x70) → Channels 0–4 → 5× MPU6050 (0x68 each)

Per-Finger Channel Mapping:

• Ch0: Thumb
• Ch1: Index
• Ch2: Middle
• Ch3: Ring
• Ch4: Pinky (hardware issue, under investigation)

Design Rationale: Per-finger IMU placement enables spatial isolation of the thumb-index interaction characteristic of Parkinsonian pill-rolling tremor, providing superior resolution over single wrist-worn sensors for fine-grained motor pattern recognition.

SPI Subsystem (Bradykinesia Detection — Pi 5 + MCP3008 integration pending; thumb-only bench characterization complete off-platform on Arduino Nano 33 BLE Sense Lite, see docs/flex-bench-characterization.md): 5× Flex Sensors → 10kΩ voltage dividers → MCP3008 ADC (CH0–CH4) → Pi SPI bus (spidev0.0)

Phone Video Subsystem (Post-Session Validation): Patient phone records encrypted session video for retrospective MediaPipe compliance checks. No real-time CV gating in the live tremor pipeline.

Pi Setup

• OS: Raspberry Pi OS 64-bit (Bookworm)
• Python: 3.11+
• I2C/SPI: Enabled via raspi-config
• Virtual Environment: source venv/bin/activate

Installed Packages

Package	Purpose
smbus2	I2C communication (IMUs)
spidev	SPI communication (MCP3008 ADC)
scipy	Butterworth filter + FFT
numpy	Sensor data processing
paho-mqtt	JSON payload publishing

Validated

• TCA9548A detected at 0x70 via i2cdetect
• MPU6050 detected at 0x68 through multiplexer
• Accelerometer data read successfully via Python (smbus2)
• 4 IMUs reading stably on channels 0-3 (Thumb, Index, Middle, Ring)
• Wearable mounting complete: PLA rings + elastic bands with hot glue on side-top bars ✅
• Multi-subject validation: 9 tests across 2 subjects (Person A: 5, Person B: 4) ✅
• Production-grade tremor discrimination: 30-895× rest-to-tremor separation ✅
• Tremor software pipeline (scripts/sensor_reader.py + scripts/dsp_pipeline.py) validated end-to-end on Pi
• Channel 4 (Pinky) stable independent read path (hardware issue + wiring crossover)
• MCP3008 ADC on SPI bus (Pi 5 host integration pending)
• Flex sensor voltage divider circuits on Pi 5 host (off-platform Arduino bench characterization complete; see docs/flex-bench-characterization.md)

Software Modules

app.py                    ← Demo web app: guided UI served on local network (port 5000)
templates/index.html      ← Web UI for app.py

scripts/
  sensor_reader.py        ← Multi-IMU polling via TCA9548A (100Hz target)
  test_imus.py            ← Per-channel probe + WHO_AM_I compatibility checks
  dsp_pipeline.py         ← Butterworth filter + FFT (4-6Hz tremor metrics)
  run_tremor_validation.py← Terminal fallback: one-command probe + capture + DSP workflow

(planned)
  transformer_inference.py← TFLite INT8 model runner (next stage)
  mqtt_publisher.py       ← JSON payload → MQTT broker (next stage)
  main.py                 ← Orchestrates staged pipeline

Current Prototype Status (Tremor Phase Complete)

Hardware

• Wearable form factor: Custom 3D-printed PLA rings (one per finger) with:
  • Bottom: Two slots for elastic band threading (adjustable tension)
  • Top: Two side bars matching sensor width for hot glue stabilization
  • Top cavity designed to fit MPU6050 sensor precisely
  • Result: Stable mounting with minimal motion artifacts
• Operational channels: 4 IMUs on channels 0-3 (Thumb, Index, Middle, Ring)
• Channel 4 issue: Hardware fault + suspected SDA/SCL crossover preventing Pinky sensor operation
• Sampling rate: 88.9-89.3 Hz sustained (4-channel), vs. 100 Hz design target
• I2C stability: Zero retries across multi-subject validation

Validation Status

• ✅ Multi-subject dataset collected: 9 tests across 2 subjects (72 per-finger DSP feature vectors)
• ✅ Tremor discrimination validated: 30-895× power increase from rest to tremor
• ✅ Clinical severity range: Light (1.4K) → Moderate (2-7K) → High (7-15K) → Severe (26K)
• ✅ Frequency accuracy: All tremor captures in 4-6 Hz Parkinsonian range
• ✅ Environmental controls documented: Contamination effects identified and logged
• ✅ Dataset ready for Transformer training: data/tremor_validation_master.csv
• 🧪 Off-platform flex bench characterization (thumb): 10 trials × 0°/30°/60° on Arduino Nano 33 BLE Sense Lite — flat-vs-bent separable (Cohen's d = 2.15 between 0° and 60°); data/flex_bench_thumb_2026-05-07.csv

Implementation Scope

• Current focus is tremor-first on the Pi 5. Flex/bradykinesia bench characterization is complete on a thumb-mounted SparkFun SEN-10264 sampled by an Arduino Nano 33 BLE Sense Lite (see docs/flex-bench-characterization.md); Pi 5 + MCP3008 multi-finger integration remains pending.
• Phone video is reserved for post-session compliance validation; no real-time camera gating in the live pipeline.
• Stable wiring topology critical: shared 3.3V and shared GND rails across Pi, TCA9548A, and all IMUs.
• scripts/test_imus.py accepts compatible WHO_AM_I responses (0x68, 0x70, 0x71) for MPU6050 clones.

Future Work

• Train Transformer/TFLite INT8 model on multi-subject validation dataset for MDS-UPDRS-aligned scoring.
• Integrate flex sensor + MCP3008 pathway on the Pi 5 host for bradykinesia and rigidity metrics (off-platform Arduino bench characterization already documented in docs/flex-bench-characterization.md).
• Add MQTT clinical payload publishing and robust session state handling.
• Add structured app/session timestamp protocol to formalize exercise semantics.
• Add security layer (payload encryption, TLS/mTLS transport, and key lifecycle controls).
• Resolve CH4 wiring/hardware issue to enable 5-channel operation.
• Evaluate and optimize sustained sampling rate toward full 100 Hz multi-channel acquisition.

Documentation

Part II — ML pipeline (CS 8674 Part II):

• part2-ml/ — Deliverable 1+ work: dataset pipeline, baseline classifiers, unified feature schema, and the week-by-week plan. Start at part2-ml/README.md; the plain-English plan is part2-ml/next-steps.md. Part II datasets live under part2-ml/data/ (gitignored).

Demo day:

• docs/DEMO_SCRIPT.md — ~5-minute spoken walkthrough for presenting at the booth
• docs/FAQ.md — answers to common questions: what are IMUs, what is bradykinesia, how was the Pi set up, what does the circuit do, and more

Setup and testing:

• docs/QUICKSTART.md — first-run setup and bring-up
• docs/testing-workflow.md — standard validation flow used in this project

Reference:

• docs/validation-results.md — latest recorded tremor-phase run outcomes (Pi 5 + IMU integrated path)
• docs/flex-bench-characterization.md — off-platform thumb flex sensor bench protocol and results (Arduino Nano 33 BLE Sense Lite)
• docs/camera-ready-edits.md — AIIoT 2026 camera-ready paper edit checklist
• docs/assessment-and-therapy-protocol.md — 3-task assessment protocol + separate therapeutic routine
• docs/mobile-web-data-contract.md — cloud payload schema for MPU tremor and flex stiffness metrics
• docs/blues-dpu-notes.md — Blues platform, DPU framing, and PD-glove integration ideas
• docs/activity5-datasets.md — measurable attributes, predictions, external datasets, and custom dataset plan
• docs/dataset-summary-2026-03-25.md — baseline dataset observations
• docs/gait-cv-pipeline.md — gait capture and CV pipeline notes

Repository Layout

pd-glove/
├── app.py                           # Demo web app (Flask server — run this for demos)
├── templates/
│   └── index.html                   # Web UI served by app.py
├── scripts/
│   ├── run_tremor_validation.py     # Terminal fallback: full rest+tremor workflow
│   ├── sensor_reader.py             # Multi-IMU polling via TCA9548A
│   ├── test_imus.py                 # Per-channel hardware probe
│   ├── dsp_pipeline.py              # Butterworth filter + FFT tremor metrics
│   └── flex_bench/
│       └── plot_bench.py            # Regenerate flex bench Fig. 5 right panel
├── data/
│   ├── tremor_validation_master.csv      # Append-only log of all assessments
│   └── flex_bench_thumb_2026-05-07.csv   # Off-platform thumb flex bench characterization (10 trials × 4 angles)
├── docs/                            # Project docs (setup, testing, specs, notes)
├── images/                          # Architecture diagrams, circuit photos
├── requirements.txt                 # Python dependencies (includes Flask)
└── README.md

How to Navigate This Repo

• Running a demo: see Running the Demo above.
• Hardware bring-up / first time setup: docs/QUICKSTART.md
• Standard testing workflow: docs/testing-workflow.md
• Current results and dataset: docs/validation-results.md
• Cloud payload schema: docs/mobile-web-data-contract.md

Running the Demo

There are two ways to run an assessment. Use the web app for demos — it gives a guided step-by-step UI in the browser. Fall back to the terminal script if Flask has issues.

Option A — Web App (primary)

app.py runs a local web server on the Pi. Open it from any browser on the same network (your laptop, tablet, etc.) — no installation needed on the client side.

ssh aqnguyen96@iotpi5.local
cd pd-glove
source venv/bin/activate
pip install -r requirements.txt   # first time only — adds Flask
python3 app.py

Then open http://iotpi5.local:5000 in your browser (or use the Pi's IP address if mDNS isn't working).

What it does:

1. Enter a Subject ID and Test Label, click Begin Assessment
2. Hardware probe runs automatically
3. 5-second countdown with prompt to hold hand still → 10s rest recording
4. 3-second countdown with prompt to shake hand → 10s tremor recording
5. DSP analysis runs, result is saved to data/tremor_validation_master.csv
6. Results page shows rest power, tremor power, severity classification, and a bar chart comparing the subject against the population average

No need to restart between subjects — click + New Assessment after each one. Use the Records tab to delete any subject's data.

Note — local hosting: The server runs only on the Pi, on your local network. Nothing is sent to the internet. Anyone on the same WiFi can reach it at port 5000 while app.py is running.

---

Option B — Terminal Script (fallback)

If the web app isn't working, use the original terminal workflow directly:

ssh aqnguyen96@iotpi5.local
cd pd-glove
source venv/bin/activate
python3 scripts/run_tremor_validation.py

The script will prompt for a Person ID and Test Name, then walk through the same hardware probe → rest capture → tremor capture → DSP → CSV append flow in the terminal.

Cleanup after a demo:

# Remove a demo subject by person ID
python3 scripts/run_tremor_validation.py --delete-person-id person_A

# More targeted: person ID + test name
python3 scripts/run_tremor_validation.py --delete-person-id person_A --delete-test-name test_one

---

Quick Start (Hardware Bring-Up)

ssh aqnguyen96@iotpi5.local
cd pd-glove
source venv/bin/activate
pip install -r requirements.txt

# Scan I2C bus (shows TCA9548A at 0x70)
sudo i2cdetect -y 1

# Test all IMUs (stable on channels 0-3)
python3 scripts/test_imus.py --channels 0,1,2,3

Testing IMU Channels

python3 scripts/test_imus.py
python3 scripts/test_imus.py --scan-only

Expected output (healthy channel):

CH0: OK WHO_AM_I=0x70 (MPU6500-class clone) Accel X = 57748

If a channel shows Remote I/O error or not found, check branch wiring continuity and verify shared 3.3V/GND rails are common across all IMUs and the multiplexer.