Detection of Situational Context From Minimal Sensor Modality of A Smartphone Using Machine Learning Algorithm

Abstract

Early detection and continuous monitoring can help reduce the complexity of treatment and recovery. For this purpose, many modern technologies are being used like smart wearable devices to make the diagnosis of different types of human diseases and automated tutoring systems. There is a vast improvement in the sector of human healthcare and education delivery using artificial intelligence (AI). For these AI algorithms, there can be high error rates if situational contexts are ignored. Currently, there is no automated approach to detect situational context. In this work, we propose a novel approach to automatically detect situational context with a smartphone context detection app using AI from minimal sensor modality. We begin the process by converting a few sensor data from the smartphone app to a multitude of axes, then determine situational context from these axes by using a machine learning algorithm. At first, we evaluated $k$-means algorithm performance on the converted data and grouped them into different clusters according to the contexts. However, the $k$-means algorithm has many challenges that negatively affect its clustering performance. For this reason, to automatically detect the situational contexts more accurately we have performed different machine learning (ML) algorithms to differentiate their characteristic parameters and attributes. To train and test ML models, 145 features were extracted from the dataset. In our case, we have used a dataset with 53,679 distinct values to evaluate the performance of different algorithms in detecting five situational contexts of the users. Experimental result shows that the accuracy of the Support Vector Machine, Random Forest, Artificial Neural Network, and Decision Tree Classifiers are 95%, 99%, 97%, and 98% respectively. The most effective classifier overall is Random Forest. This preliminary work shows the feasibility of detecting situational context automatically from a few sensor data collected from the smartphone app by converting the sensor data to multiple axes and applying a machine learning algorithm.