Intelligent Sensor Systems for Healthcare: A Case Study of Pressure Ulcer TITOLO Prevention and Treatment TESI Rui (April) Dai Assistant Professor Department of Computer Science North Dakota State University 1
Sensor Systems for Healthcare Low power and implanted medical sensors Smartphones Computer networking Real-time, long-term monitoring of physical activity, physiological state, etc. 2
Background Center for Assistive Technology and Environmental Access Enhance the health and function of persons with disabilities Multidisciplinary research team Shepherd Center Medical treatment, research, and rehabilitation for people with spinal cord injury and brain injury 3
Pressure Ulcer Pressure ulcer (PU): a type of skin injury Primary biomechanical factor for PU development: internal tissue deformation resulting from external interface pressure. Persons with spinal cord injuries (SCI) sit in wheelchairs for many hours per day -> They are at high risk of developing PUs in their buttock areas A major problem for people with SCI More than 50% of the individuals with SCI experience pressure ulcers during their lifetimes 4
Contributions Pressure ulcer prevention A wheelchair pressure relief monitoring system Robust Long-term monitoring Pressure ulcer treatment Wound measurement system Handheld device Non-contact High accuracy and reliability 5
Contributions Pressure ulcer prevention A wheelchair pressure relief monitoring system Robust Long-term monitoring Pressure ulcer treatment Wound measurement system Handheld device Non-contact High accuracy and reliability 6
Pressure Ulcer Prevention Strategies Design and select pressure-relieving cushions that can distribute the body weight evenly and over a large area Example of a cushion: air bags + foam base Perform regular pressure reliefs (PR) 7
Pressure Relief Monitoring (PRM) System R. Dai, S. Sonenblum, and S. Sprigle, A Robust Wheelchair Pressure Relief Monitoring System, IEEE EMBC, Aug. 2012. Study the relationship between pressure relief behavior and pressure ulcer development Provide real-time instructions to remind wheelchair users when pressure relief is necessary Monitor pressure relief behaviors for wheelchair users Achieve long-term monitoring Automatically detect pressure relief events (starting times and durations) 8
Interface Pressure Measurement Existing monitoring systems Interface pressure mat (IPM): placed on top of the wheelchair cushion to perform fine-grained measurement Commercially available IPM: FSA (Force Sensing Array) IPM with 16*16 pressure sensors 9
Limitations for IPM-Based Monitoring Interfere with the performance of wheelchair cushion Increase interface pressures Make the surface of the cushion slippery, resulting in sliding and a poor seated posture High power consumption and large data volume Works for only 1 hour on battery The costs for processing and memory are not negligible None of the existing IPM based systems are used clinically for long-term monitoring 10
Pressure Relief Monitor (PRM): Equipment PRM sensors: 8 custom pressure sensors packaged in two mats, each containing 4 individual 2 x2 sensors and sealed within a robust polyethylene cover Placed under the wheelchair cushion do not interfere with cushion performance 11
Pressure Relief Monitor: Equipment Sensor placement The tissue near the person s sitting bones experiences the greatest loads during sitting, and therefore is at high risk of developing pressure ulcers. Place the sensors to relate to the pressures around a person s sitting bones. Based on tests on subjects with different body sizes, decide the best location to place the sensors. L = 15cm H = 7cm 12
Pressure Relief Monitor: Power Supply The system can work at a sampling rate of 1Hz on two AA batteries for several weeks Battery, interface circuit, and data loggers 13
Methodology Pressure relief is achieved when the pressure on the interface between the buttock area and the cushion is reduced to a certain extent How is the interface pressure transmitted to the bottom of the cushion? It varies based on the specific materials and construction of the cushion Determining pressure reliefs from the sensors beneath the cushion is not obvious 14
Methodology Supervised learning Individualized classifier The size, weight, and capability to perform pressure reliefs can be very diverse The materials of wheelchair cushions vary widely 15
Methodology 16
Training Protocol Short-term training in the beginning of a test Upright Full front lean Medium front lean Small front lean Upright Push up Upright Upright Full left lean Full right lean Medium left lean Medium right lean Small left lean Small right lean Upright Upright Push up Push up With each position held for 20 seconds, the training protocol lasts for 10 minutes 17
Data Labeling Interface pressure Measured by an interface pressure mat (IPM) Used to label the training data as pressure relief or upright sitting (no pressure relief) Calculate peak pressure index (PPI) Defined as the highest recorded pressure values within a 9-10 cm 2 area under the sitting bones Identify left and right sitting bone regions with manual palpation 18
Data Labeling Differentiate the PPI values for pressure relief and upright sitting Given the variance in PPI values across subjects, normalize the raw values of PPI Find the average of PPI for each upright segment in the training sequence: PPI upright_avg Calculate a normalized PPI by PPI norm =PPI/PPI upright_avg 19
Data Labeling Studied the distribution of PPI norm Upright sitting Intended pressure reliefs Decide a cut-off threshold: T=0.85 P(PPI intended_pressure_reliefs <T) = 98% P(PPI upright >T) = 95% Define that pressure relief happens when one side of the normalized PPI is smaller than the cut-off threshold i.e., when PPI norm_left <T or PPI norm_right <T. 20
Pressure Relief Behavior Classification Stationary segments and non-stationary segments Non-stationary segments posture change/noise Only data from the stationary segments are sent to the classifier Find stationary segments based on measurements of the 8 PRM sensors Apply a moving window of 5 seconds, and the data inside a window is stationary if for each of the 8 sensors measurements, the deviation are within a threshold Extract features from the stationary segments 21
Feature Extraction Feature 1: Center of pressure Calculated from the pressure and location of each sensor Feature 2: Peak pressure The largest pressure of the 8 sensors 22
Feature Extraction Feature 3: FrontBackRatio Ratio of the sum of pressure in the front to that in the back Feature 4: MinMaxFrontBack Find the maximum pressures of sensors placed in the back and those placed in the front, and find the minimum of the two Feature 5: MinMaxLeftRight Find the maximum pressures of sensors placed at the left side and those at the right side, and find the minimum of the two 23
Classification K-nearest neighbor classification Sensitivity The percentage of pressure relief points which are correctly identified as pressure reliefs Specificity The percentage of no pressure relief points that are correctly identified 24
Performance Evaluation: Inter-Cushion Test On the same, able-bodied subject Four commonly used cushions with diverse materials 10 minutes training protocol+40 minutes free sitting+10 minutes training protocol First 10 minutes data for training, test on the remaining 50 minutes Accuracy (%) Foam (Standard foam) Matrx (Memory foam) Cushion Jay (Gel with foam base) Roho (Adjustable air cushion) Sensitivity 85.3 94.3 93.3 86.7 Specificity 97.9 82.1 92.1 89.9 25
Performance Evaluation: Inter-Person Test 3 participants with spinal cord injuries who use wheelchairs as their primary mobility devices (*The protocol was approved by the local Institutional Review Boards) Test on their own wheelchairs and cushions Each subject performed the training protocol in the beginning of the test, and repeated the protocol one week later Average sensitivity: 91%, average specificity: 89% Subject No. 1 2 3 Gender Male Female Male Age 25 54 24 Weight(lbs) 160 120 180 Cushion Roho Matrx Jay Accuracy Subject No. (%) 1 2 3 Sensitivity 83.0 92.2 98.0 Specificity 95.3 85.7 84.9 26
Performance Evaluation Error analysis Small movements Placing sensors beneath the cushion is not as sensitive as measuring interface pressure directly For some cushions, the change in interface pressure is transmitted to the underneath sensors with a few seconds delay 27
Contributions Pressure ulcer prevention A wheelchair pressure relief monitoring system Robust Long-term monitoring Pressure ulcer treatment Wound measurement system Handheld device Non-contact High accuracy and reliability 28
Wound Measurement Monitoring wound size An essential component to the assessment and treatment of pressure ulcers and other chronic wounds Tracking changes or improvements of a wound Area measurement Depth measurement (maximum depth) Goals Accuracy Reliability 29
Wound Measurement Ruler-based method: most widely used Need to contact the wound Inaccurate for irregular shaped wounds Current vision-based techniques 3-D construction: taking two or more photographs of the same wound from different angles Expensive and not portable Need a new wound measurement system Non-contact Portable Low cost 30
Handheld Non-Contact Wound Measurement Device Smart phone with camera sensor Lasers with known positions to assist wound measurement (a) Side view (b) Back view (c) Device in use 31
Wound Measurement Given Camera s sensing parameters (calibration) Position of lasers with respect to the camera Locations of laser points in the captured image From the camera s projection geometry, calculate Skew angle: to perform correction Scaling ratio: real world size of a pixel 32
Wound Area Measurement Skew correction Perform edge detection on the image Calculate the number of pixels inside the wound Calculate the actual area of the wound 33
Wound Depth Measurement Add a line focused laser on the device The projected line laser on the wound surface reflects depth information (a) Line focused laser installed on the device (b) Laser projection (c) Captured image for depth measurement 34
Wound Depth Measurement Let a user click four points on the projected laser line Two outside the wound Two on the edge of the wound Detect the laser line outside the wound, and segments of laser line inside the wound. Device Calculate depth based on the skew angle, scaling ratio, and detected laser lines r D d tan( ) cos Wound plane Camera s sensing direction r D r D/ cos d Laser plane Target plane 35
Performance Area measurement Error rate: 2%-6% Under skew angle: 0 15 degrees Depth measurement Error rate: 8% Under skew angle: 0-15 degrees Reliability Inter-user variance 10% Enhancements based on clinical feedback User-friendly graphical interface Human-computer interactive correction on measured results 36
Research Summary Sensor-based systems to support clinical research Pressure relief monitoring for pressure ulcer prevention Wound measurement for pressure ulcer treatment * Local processing With communication and networking Remote monitoring Remote control Integration of information from different sensors/people 37
Ongoing Work Wireless body area networks for rehabilitation Objectives: Real-time and reliable computation and communication to derive meaningful physiological parameters Maximize the lifetime of sensors and aggregators 38
Thank you. 39