Introduction
As the global population ages, the demand for technologies that support independent living has reached a critical point. While many elderly individuals prefer to remain in their own homes, safety concerns often necessitate a move to assisted living facilities. Current technological solutions, primarily focused on fall detection, frequently fall short due to privacy concerns, high false-alarm rates, and a lack of diagnostic depth. This paper presents an intelligent monitoring system designed to address these gaps by utilizing wearable inertial sensors and advanced signal processing to monitor both acute events and long-term health trends.
The Limitations of Current Monitoring Systems
Existing monitoring technologies generally fall into two categories: video-based systems and manual alarm buttons. Both have significant drawbacks:
- Privacy and Intrusion: Video-based monitoring is often perceived as intrusive, leading to low adoption rates among the elderly.
- User Dependency: Manual alarms require the user to be conscious and physically capable of pressing a button during an emergency, which is often not the case during a severe fall or medical episode.
- False Alarms: Simple threshold-based sensors often fail to distinguish between high-intensity daily activities (such as sitting down quickly) and actual falls.
- Diagnostic Gap: Most systems are reactive, detecting an event after it occurs rather than identifying the physiological deterioration that leads to such events.
Proposed Methodology
The proposed system employs a two-step classification process to provide a comprehensive view of the user's status. By moving beyond simple event detection, the system acts as a continuous health monitoring tool.
Two-Step Classification
- Activity Classification: The system first categorizes movement into five primary types: fall, unconscious fall, walking, standing/sitting, and lying down.
- Health State Classification: If the system identifies the activity as "walking," it triggers a secondary analysis of the gait. This step distinguishes between five distinct health states: healthy, hemiplegia (often resulting from a stroke), Parkinson’s disease, leg pain, and back pain.
Multidimensional Dynamic Time Warping (DTW)
The technical core of the system is the use of multidimensional Dynamic Time Warping (DTW) as a similarity measure within a k-nearest neighbor (k-NN) classifier. Unlike standard DTW, which typically compares single-variable sequences, the multidimensional approach analyzes time-series data from multiple sensors (accelerometers and gyroscopes) simultaneously. This allows the system to compare the "shape" of movement across all axes, making it highly robust to variations in an individual's walking speed or stride duration.
Experimental Evaluation
The system was evaluated using a dataset of 256 recordings involving both healthy subjects and individuals with specific health conditions (e.g., Parkinson’s, hemiplegia). To ensure the classifier's robustness, each subject was recorded multiple times performing movements at varying speeds.
Initially developed using a high-precision, 12-tag infrared motion capture system, the methodology was successfully adapted for wearable inertial sensors. This transition is crucial for real-world application, as it enables continuous monitoring both indoors and outdoors without the need for expensive, fixed infrastructure.
Key Results and Insights
The multidimensional DTW approach achieved classification accuracies exceeding 96% for both activity detection and health-state recognition. Several key insights emerged from the study:
- Gait as a Diagnostic Tool: By identifying the specific "signature" of conditions like Parkinson’s or stroke-related gait, the system provides actionable data for physicians, potentially leading to earlier diagnoses.
- Preventative Monitoring: By detecting the onset of instability caused by leg or back pain, the system can alert caregivers to an increased fall risk before an injury occurs.
- Algorithmic Flexibility: The shift from manual feature engineering (e.g., calculating specific joint angles) to DTW-based analysis of raw sensor data makes the system more adaptable to different hardware configurations and reduces the need for complex pre-processing.
Conclusion and Future Work
The research demonstrates that high-accuracy health monitoring is possible using relatively inexpensive wearable sensors. The system is currently being commercialized to provide a service for elderly individuals seeking to maintain their independence.
Future work will focus on further simplifying the hardware requirements, with the goal of achieving similar accuracy using a single sensor, such as a smartphone. Additionally, the research team plans to expand the system's scope to include cardiac monitoring by integrating inertial data with physiological sensors, providing a more holistic view of elderly health.
Summary updated