Continuous physical health monitoring and classification for distress detection Daniela Cr˘ aciun 1 June 28, 2017 1 Babes ¸-Bolyai University, Cluj-Napoca, Romania 

Introduction 

Context: Health Monitoring Falling can be very dangerous to people with mobility issues. Even if no injuries occur, it can have long term effects. ⇒ People at high risk of falling may also usually suffer of other diseases, such as: • Muscular dystrophy • Parkinson’s disease • Stroke • Arrhythmia Continuous health monitoring is ideal when dealing with mobility issues, providing essential information to a caregiver (for effective intervention in case of emergency) 1/13 

Fall related statistics: Type of injuries 7 % Concussion or brain injury 21 % Contusion 12 % Open wound 6 % Sprain 46 % Fracture 8 % Other Figure 1: Type of injuries people report after a fall 2/13 

Complications ⇒ more old/disabled people in need of care (due population growth) ⇒ existing systems have issues: • high price • hard to use by people with no medical training • uncomfortable or hard to use • rely on the care receiver to alert caregivers ⇒ heart rate could be a indication to a person’s state, but fluctuation ca be induce either by mental or physical activity 3/13 

Premise Question: how to obtain a more accessible system for complete monitoring? An idea: correlate heart rate with the person’s movement to provide an all-around continuous evaluation of health (fall detection, activity level and heart status information) with consumer-level devices 4/13 

Design and Methodology 

Structural overview Composed of multiple elements, referred to as units: • Sensor unit is composed of: • Movement unit, which uses an IMU • Heart rate unit, uses a pulse sensing mechanism • Router unit • cleans and stores data • synchronizes information • ensures communication between all the components • provides client-side interface • relies on Python (Flask framework) as server-side language and basic Bootstrap on the front-end • Prediction unit • uses random forest algorithm for activity level prediction • determines heart status, correlated with activity level • Python library scikit-learn used for support 5/13 

Structural overview Figure 2: Monitoring system structure 6/13 

Web-based interface • information is displayed to the caregiver using a web-based interface • the caregiver provides information about the care receiver and connects the two sensor-based devices using provided forms Figure 3: Web panel interface 7/13 

Theoretical details 

Random forests Random forest is an ensemble method (meaning that it contains a set of predictive models) in which: • individual predictions from multiple decision trees are combined using voting H(x) = labelmax occurence(h1(x),h2(x),...,hn(x)) where the n is the number of independent classifiers and h(x) represents an individual prediction. • each tree is generated based on a subset from the entire dataset and m attributes is given to each tree to predict on (m is constant during the prediction process) 8/13 

Random forests: Structure Figure 4: Random forest classifier structure 9/13 

Data and Results 

Dataset choice The application uses the Physical Activity Monitoring dataset (consisting of 54 attributes), from which a subset was used. Only one of the 3 original IMU is considered (trunk one yields best results). Intensity level Activities Light Alert walking, Descending stairs, Vacuum cleaning, General house cleaning Moderate Lying, Sitting, Standing, Walking, Watching TV, Computer work, Driving a car, Ironing, Folding laundry High Running, Cycling, Ascending stairs Table 1: Original dataset activities mapping by energy expenditure Heart rate column is sparse, so it is replaced with 0 in the cases it is missing. Finally, 11 attributes remain: acceleration, gyroscope and magnetometer data (3 columns each), heart rate and activity level 10/13 

Results: comparing with related work Related activity level prediction systems Activity classifier Accuracy K-Nearest Neighbors 1 70% Rotation Forest 1 94% Neural Networks 1 89% Neural Networks 2 93% Random Forests (this implementation) 91% Table 2: Comparisons with related work 1H. Xu, J. Liu, H. Hu, and Y. Zhang. Wearable Sensor-Based Human Activity Recognition Method with Multi-Features Extracted from Hilbert-Huang Transform. Sensors, 2016 2M. Arif and A. Kattan. Physical activities monitoring using wearable acceleration sensors attached to the body. PloS one, 2015. 11/13 

Conclusions and Future Work 

Summary Advantages of proposal • intuitive web-panel based interaction for caregiver • consumer-level devices used • easy to extend and reusable 12/13 

Next steps • use a lighter sensor unit • could be more useful in a portable way • offer wider range of vital signs tracking • monitor more than one user at once 13/13 

Thank you! 13/13 

Questions? 13/13