A Low-Cost Wearable Gait Analyzer Using IMU Sensors and Machine Learning for Early Parkinson’s Tremor Prediction: An Electrical Engineering Perspective

A Low-Cost Wearable Gait Analyzer Using IMU Sensors and Machine Learning for Early Parkinson’s Tremor Prediction: An Electrical Engineering Perspective

Akshita Murali

Department of Electrical and Electronics Engineering Amrita Vishwavidyapeetham, Coimbatore, India

Abstract – Parkinson's disease (PD) is a progressive neurodegenerative disorder affecting motor control, with early symptoms including subtle gait abnormalities and resting tremors. Early detection is crucial for timely intervention and disease management. This paper presents a low-cost, wearable gait analysis system from an electrical engineering perspective, utilizing an Inertial Measurement Unit (IMU) sensor with comprehensive signal conditioning and power management circuits, integrated with classical machine learning algorithms for early tremor prediction. The system employs adaptive power management techniques to ensure extended battery life while maintaining signal fidelity for accurate feature extraction. Designed with cost-effective commercial-off-the-shelf (COTS) components, the device incorporates proper signal conditioning, noise filtering, and wireless communication modules. Experimental validation shows classification accuracy exceeding 90% while achieving 48 hours of continuous operation on a single charge, demonstrating the system's efficacy as a non-invasive, power-efficient solution for early PD screening and long-term monitoring.

Keywords: Parkinson's disease, IMU sensors, Machine Learning, Gait analysis, Tremor detection, Wearable device, Power management, Signal conditioning, Embedded systems, Biomedical instrumentation

1. INTRODUCTION

   Parkinson's disease remains a significant neurological challenge, with early diagnosis complicated by subtle motor symptom onset. From an electrical engineering standpoint, developing wearable biomedical devices for PD monitoring requires addressing critical challenges in signal acquisition integrity, power management, and embedded system design. Traditional clinical assessments lack continuous monitoring capability and are subject to inter-rater variability. This paper presents an integrated approach combining analog signal conditioning, digital signal processing, and machine learning classification to create a wearable system that bridges the gap between clinical accuracy and everyday usability. The system emphasizes electrical design principles to ensure reliable operation while maintaining low cost and extended battery life.

2. PROBLEM STATEMENT

   Existing wearable PD monitoring systems often compromise either signal quality or power efficiency, leading to either short operational lifespans or inaccurate measurements. Many commercially available systems use unoptimized sensor interfaces that introduce noise artifacts, while power management is frequently an afterthought rather than a design constraint. There is a critical need for an electrically optimized system that addresses:

   1. Signal integrity through proper sensor interfacing and conditioning

   2. Power efficiency through intelligent sleep modes and power gating

   3. Cost-effectiveness through careful component selection

   4. Reliability through robust circuit design and error handling

3. Objectives

   1. To design and prototype an electrically optimized wearable system with:

      • Proper IMU sensor interfacing circuits

      • Adaptive power management with multiple sleep states

      • Efficient wireless data transmission

      • Signal conditioning for noise reduction

   2. To develop a hybrid analog-digital signal processing pipeline that maximizes feature extraction accuracy while minimizing computational load.

   3. To implement energy-aware machine learning algorithms suitable for microcontroller deployment.

   4. To validate the system's electrical performance through power measurements, signal-to-noise ratio (SNR) analysis, and battery life testing.

4. Literature Review: Electrical Design Perspectives

   Previous research in wearable PD monitoring has often focused on algorithm development while overlooking electrical design considerations. Studies have shown that improper sensor mounting and inadequate signal conditioning can introduce motion artifacts that significantly degrade classification accuracy. Power management strategies in existing systems typically employ simple sleep- wake cycles without considering the specific power profiles of PD monitoring tasks. Recent advances in ultra-low-power microcontrollers, energy harvesting techniques, and adaptive sampling algorithms provide opportunities for system optimization. This work distinguishes itself by integrating electrical design principles with machine learning, creating a system where hardware and software are co-optimized for PD monitoring.

5. PROPOSED SYSTEM: ELECTRICAL ARCHITECTURE

   1. Hardware Subsystem Design

      1. Sensor Interface Circuit

         IMU (MPU6050) Analog Front-End Anti-aliasing Filter ADC (16-bit)

         I²C Interface Microcontroller

         • Analog Front-End: Instrumentation amplifier with gain=10 for weak tremor signals

         • Anti-aliasing Filter: 4th-order Butterworth low-pass filter (fc = 25 Hz)

         • ADC Selection: Integrated 16-bit SAR ADC ( for better noise performance)

      2. Power Management Unit (PMU)

         Li-ion Battery (3.7V, 1000mAh) Buck-Boost Converter (TPS63020)

         Power Distribution Network

         [1.8V] IMU Sensor [3.3V] MCU Core [3.3V] Wireless Module

         • Multiple Voltage Domains: Separate LDOs for analog and digital sections

         • Dynamic Voltage Scaling: Adjusts MCU frequency based on processing load

         • Power Gating: Individual enable/disable for sensor, wireless module, and processing core

      3. Microcontroller Selection Criteria

         • ESP32-S3: Dual-core, ultra-low-power modes (10A in deep sleep)

         • Integrated Features: Hardware accelerators for FFT and matrix operations

         • Peripheral Optimization: Direct memory access (DMA) for sensor data collection without CPU intervention

   2. System Architecture Block Diagram

   3. Electrical Specifications

      Parameter	Specification	Design Consideration
      Operating Voltage	3.3V ±5%	Optimized for battery discharge curve
      Current Consumption	Active: 45mA, Sleep: 12A	Enables 48+ hours continuous operation
      ADC Resolution	16-bit	Adequate for 0.01g tremor detection
      Sampling Rate	Configurable: 25-100 Hz	p>Adaptive based on activity detection
      Wireless Protocol	Bluetooth 5.0 Low Energy	Balance between range and power
      Battery Life	>48 hours continuous monitoring	Achieved through duty cycling (5% active)
      Signal SNR	>40 dB after conditioning	Ensures reliable feature extraction

   4. Cost Analysis (Indian Market)

      Component	Model/Specification	Quantity	Unit Price ()	Total Cost ()
      Microcontroller	ESP32-S3 (Development Board)	1	350	350
      IMU Sensor	MPU6050 (6-axis)	1	120	120
      Li-ion Battery	1000mAh, 3.7V	1	150	150
      Battery Management IC	TP4056	1	25	25
      Buck-Boost Converter	TPS63020	1	85	85
      Operational Amplifiers	MCP6002 (Dual Op-Amp)	2	40	80
      Passive Components	R, C, L (SMD packages)	1 set	100	100
      PCB Fabrication	2-layer, FR4	1	200	200
      Enclosure	3D Printed PLA	1	50	50
      Strap & Fasteners	Adjustable Velcro	1 set	60	60
      Subtotal (Prototype Cost)				1,220
      Estimated Mass Production	(1000 units, including assembly)			750-850

      Total Prototype Cost: 1,220

      Estimated Mass Production Cost: 750-850 per unit

6. METHODOLOGY: ELECTRICAL IMPLEMENTATION

   1. Signal Acquisition and Conditioning

      1. Analog Signal Path Design

         Raw IMU Instrumentation Amp (G=10) 1st Stage LPF (fc=50Hz)

         2nd Stage Active Filter Programmable Gain Amp ADC Input

         • Noise Analysis: Calculated input-referred noise = 150V RMS

         • Common Mode Rejection: >80 dB at 60 Hz (power line rejection)

         • Dynamic Range: 0-4g with 0.01g resolution (sufficient for tremor detection)

      2. Digital Signal Processing Pipeline

         ADC Output Moving Average Filter IIR Notch Filter (50/60 Hz)

         Bandpass Filter (0.5-12 Hz) Feature Extraction Classification

   2. Power Management Strategy

      1. Multi-level Sleep Architecture

      2. Adaptive Sampling Algorithm

         text

         if (activity_detected == True): sampling_rate = 100 Hz enable_all_peripherals()

         elif (suspected_tremor == True): sampling_rate = 50 Hz enable_IMU_only()

         else:

         sampling_rate = 10 Hz enter_light_sleep_between_samples()

   3. Machine Learning Implementation for Embedded Systems

      1. Feature Selection for Power Efficiency

         Features selected based on computational complexity and discriminative power:

         • Low-compute features: Mean, variance, zero-crossing rate

         • Medium-compute features: FFT-based spectral features (hardware accelerated)

         • Avoided features: Wavelet transforms (computationally expensive)

      2. Model Optimization for Microcontrollers

         • Quantization: 8-bit integer arithmetic for inference

         • Pruning: Removed 40% of least important Random Forest features

         • Memory optimization: Feature calculation in-place to minimize RAM usage

   4. PCB Design Considerations

      1. Layer Stackup: 4-layer board with dedicated ground plane

      2. Component Placement: Separated analog and digital sections

      3. Routing: Minimized high-speed trace lengths, proper impedance matching

      4. Shielding: EMI shielding for sensor and wireless sections

      5. Test Points: Included for debugging and performance measurement

   5. Dataset Description and Validation Methodology

      The machine learning model was trained and evaluated using a publicly available Parkinsons disease gait and tremor dataset. The dataset consists of inertial sensor recordings collected from Parkinsons disease patients and healthy control subjects during controlled walking and resting tasks. Each sample includes tri-axial accelerometer and gyroscope measurements, recorded at sampling rates comparable to those used in the proposed wearable system.The recorded signals were segmented into fixed-length windows and labeled according to subject condition. A 70:30 traintest split was employed along with 5-fold cross-validation to ensure robustness and to reduce overfitting. Model performance was evaluated using classification accuracy, power consumption per inference, and memory usage on the target embedded platform.The proposed system is intended for screening and long-term monitoring and is not designed to replace clinical diagnosis.

7. EXPERIMENTAL RESULTS: ELECTRICAL PERFORMANCE

   1. Power Consumption Analysis

      Operation Mode	Current	Duration	Energy per Cycle
      Active Processing	45 mA	200 ms	9 mJ
      Data Transmission	28 mA	50 ms	1.4 mJ
      Light Sleep	850 A	1.8 s	1.53 mJ
      Deep Sleep	12 A	Variable	Minimal
      Average	~2.1 mA	Continuous	Projected: 54 hours

   2. Signal Quality Metrics

      • Signal-to-Noise Ratio: 42.3 dB (after conditioning, 18.7 dB raw)

      • Effective Number of Bits (ENOB): 13.2 bits (from 16-bit ADC)

      • Harmonic Distortion: <1% THD at 5 Hz (tremor frequency range)

      • Crosstalk Between Axes: <-60 dB

   3. Classification Performance vs. Power Consumption

      Feature Set	Accuracy	Power per Inference	Memory Usage
      Time-domain only	86.2%	2.1 mJ	2.1 KB
      Frequency-domain only	88.7%	3.8 mJ	3.5 KB
      Combined (proposed)	92.3%	4.2 mJ	4.8 KB
      Deep Learning (baseline)	94.1%	82.5 mJ	156 KB

   4. Thermal Performance

      • Maximum temperature rise: 3.2°C above ambient during continuous operation.

      • No thermal throttling required.

   5. Cost-Performance Comparison

      System	Accuracy	Battery Life	Cost ()	Cost per 1% Accuracy ()
      Proposed System	92.3%	54 hours	1,220	13.22
      Commercial Research Device [1]	94.1%	24 hours	15,000	159.40
      Smartphone-only Solution [2]	84.5%	N/A	0*	0
      Clinical Motion Capture	96.8%	N/A	8,00,000+	8,264.46

      *Assumes user already owns smartphone

8. DISCUSSION: ELECTRICAL ENGINEERING CONTRIBUTIONS

   1. Innovations in Power Management

      The proposed adaptive power management system extends battery life by 300% compared to conventional always-on designs. By implementing a state machine that transitions between power modes based on detected activity, the system maintains responsiveness while minimizing energy consumption. The total power budget of 2.1 mA average current is significantly lower than comparable systems reported in literature (typically 5-10 mA).

   2. Signal Integrity Enhancements

      The custom analog front-end improved SNR by 23.6 dB compared to direct IMU-to-MCU connections. This enhancement proved critical for detecting early-stage tremors with amplitudes as low as 0.02g. The effective resolution of 13.2 bits from a 16-bit ADC indicates minimal noise contamination in the signal chain.

   3. Cost-Performance Optimization

      At 1,220 per prototype unit (750-850 in mass production), the system achieves a cost-to-performance ratio superior to existing solutions. The cost per 1% accuracy metric demonstrates that the proposed system provides excellent value, being 12 times more cost-effective than commercial research devices and 625 times more affordable than clinical motion capture systems.

   4. Limitations and Mitigations

      • Battery aging: Implemented coulomb counting for state-of-charge estimation

      • Wireless interference: Frequency hopping and retry mechanisms in BLE stack

      • Motion artifacts: Adaptive filtering based on activity classification

      • Environmental variations: Temperature compensation in sensor calibration

9. FUTURE WORK: ELECTRICAL ENHANCEMENTS

   1. Energy Harvesting Integration: Incorporate piezoelectric (150 additional cost) or thermoelectric harvesting for self- sustaining operation

   2. Advanced Power Management IC: Custom ASIC design integrating PMU, sensor interface, and preprocessing

      (estimated cost reduction: 200/unit in volume)

   3. Multi-sensor Fusion: Add EMG sensors (300 additional) with isolated front-ends for comprehensive motor assessment

   4. Wireless Power Transfer: Qi-standard charging (250 additional) for improved user convenience

   5. FPGA Acceleration: Low-power FPGA for real-time feature extraction (increases cost by 500 but reduces power by

      30%)

   6. Biocompatible Encapsulation: Medical-grade silicone coating (100 additional) for long-term wearability

   Estimated advanced version cost: 1,500-1,800 per unit

10. CONCLUSION

    This paper presents an electrically optimized wearable system for early Parkinson's disease detection that balances signal fidelity, power efficiency, and cost-effectiveness. By applying electrical engineering principles to system designfrom analog signal conditioning to power management and embedded ML implementationwe have developed a practical solution suitable for long-

    term home monitoring. The system achieves 92.3% classification accuracy while operating for over 48 hours on a single charge at a prototype cost of 1,220 (750-850 in mass production), demonstrating the feasibility of low-cost, high-performance wearable medical devices. The electrical design optimizations result in a system that is 12 times more cost-effective than commercial alternatives while maintaining comparable performance. Future work will focus on miniaturization, additional sensor modalities, and clinical validation with larger patient cohorts.

11. REFERENCES

    1. M. R. Patel et al., "Wearable Sensors for Parkinson's Disease: A Review," IEEE Sensors Journal, vol. 18, no. 20, 2018.

    2. A. S. Tabesh et al., "Low-Power Signal Conditioning Circuits for Wearable Sensors," IEEE Transactions on Biomedical Circuits and Systems, vol. 14, no. 2, 2020.

    3. J. Yoo et al., "A 5.2W 0.0046mm² Analog Front-End for Wearable EEG Systems," ISSCC Digest of Technical Papers, 2019.

    4. P. De Lima et al., "Machine learning for tremor detection in Parkinson's patients," Biomedical Signal Processing, 2019.

    5. R. J. M. Vullings, "Adaptive Power Management for Wearable Medical Devices," IEEE Transactions on Biomedical Engineering, vol. 67, no. 3, 2020.

    6. K. Y. Chan et al., "Ultra-Low-Power Design Techniques for Biomedical Applications," IEEE Circuits and Systems Magazine, vol. 20, no. 2, 2020.

    7. B. R. Bloem et al., "The hallmarks of Parkinson's disease," FEBS Journal, 2021.

    8. ESP32 Series Datasheet, Espressif Systems, 2022.

    9. MPU-6000/MPU-6050 Product Specification, TDK InvenSense, 2020.

    10. Texas Instruments, "Power Management Guide for Wearable Applications," Application Report, 2021.

    11. S. K. Gupta, "Cost Analysis of Wearable Medical Devices in Indian Context," Indian Journal of Medical Electronics, vol. 24, no. 3, 2022.

    12. R. Sharma et al., "Affordable Healthcare Technology for Developing Nations," IEEE Transactions on Healthcare Systems Engineering, 2021.

12. APPENDIX: CIRCUIT SCHEMATICS AND LAYOUT

1. Bill of Materials (Indian Pricing)

   Part	Value/Type	Package	Qty	Price ()	Supplier
   ESP32-S3	Dev Board	–	1	350	Robu.in
   MPU6050	6-axis IMU	QFN-24	1	120	Element14
   Li-ion Battery	1000mAh	402030	1	150	LG India
   TP4056	Charger IC	SOP-8	1	25	Texas Instruments
   TPS63020	Buck-Boost	QFN-10	1	85	Texas Instruments
   MCP6002	Dual Op-Amp	SOIC-8	2	80	Microchip
   Resistors	0603 SMD	0603	30	30	local
   Capacitors	0603 SMD	0603	25	40	local
   Inductors	4.7H	0805	3	30	TDK
   PCB	50x50mm, 2-layer	FR4	1	200	PCBWay India
   Enclosure	3D Printed	Custom	1	50	Local print
   Total				1,220

2. PCB Design Guidelines Followed

   1. Component Placement: Analog section isolated from digital

   2. Power Traces: 20 mil width for main power lines

   3. Ground Plane: Continuous on bottom layer

   4. Decoupling: 100nF capacitors within 2mm of each IC

   5. Test Points: Included for all critical signals

3. Assembly Cost Breakdown (for Mass Production)