Wearable Based-Sensor Fall Detection System Using Machine Learning Algorithm

• Published in: 2021 8th International Conference on Computer and Communication Engineering (ICCCE) pp. 53 - 57
• Main Authors: Ishak, Anis Nadia, Habaebi, Mohamed Hadi, Yusoff, Siti Hajar, Islam, Md. Rafiqul
• Format: Conference Proceeding
• Language: English
• Published: IEEE 22.06.2021
• Subjects: AndroSensor; fall; Machine learning; Machine learning algorithms; MATLAB; Real-time systems; Senior citizens; Sensitivity; Software; Support vector machines; SVM
• DOI: 10.1109/ICCCE50029.2021.9467195

As nations develop and prosper economically, their population ages longer and requires extra healthcare. Falls are known to be the second major factor of deaths in elderly by accidental or unwarranted injuries. When a fall occurs, lack of immediate help or action is the main problem, especially when bleeding is involved, as fall-related injuries are a life-threatening for many people. To prevent such kinds of deadly scenarios, a reliable fall detection system must be developed to help many lives. In this project, a wearable sensor-based fall detection system using a machine-learning algorithm had been developed. An application called 'AndroSensor' on a smartphone, that retrieves real-time data from accelerometer, gyroscope and gravity sensors, is used as the input signals. The phone is placed at the most accurate position that had been done by past research which is waist position. When a fall event occurs, the real-time data is collected and placed in a *.CSV file. Then, a Machine Learning Algorithm (MLA) is used to train and test the data before a classifier is used to classify the new incoming dataset. The fall event behaviour classification classes are sleep, walk, sit, front fall, back fall, side fall, etc. MATLAB software is currently used to analyse and visualize the data too. The MLA detects fall with efficient sensitivity (SP), specificity (SP), and accuracy. An accuracy of 100% is achieved with the Support Vector Machine (SVM) classifier compared to other classifiers has been confirmed by many past research. However, other classifiers like Decision Tree and kNN had 100% accuracy too. This means that the proposed system achieved its goals. As for future work, the plan is to convert the code to an app to run on the smartphone so it can be commercialized.