Physical activity forecasting with time series data using Android smartphone

Physical Activity Forecasting (PAF) has several important use cases such as the real-time identification of abnormal behavior, diagnosis of certain diseases, and adherence to physical activity (PA) prescriptions. In this paper, we describe a forecasting algorithm for future human PA using smartphones. Specifically, we first categorize PAs into ‘still’, ‘tilting’, ‘walking’, ‘running’, ‘in vehicle’, and ‘on bicycle’. These categories are detected by our Android smartphone application and stored in its storage. We also measure the step count and calorie consumption of our users. We then utilize Bidirectional Long-Short-Term Memory (BiLSTM) for forecasting future human PA. The collected data of PA, step, and calorie counts are uploaded to the cloud storage, and trained by the BiLSTM. Our approach forecasts the users’ PA with its duration, step, and calorie counts at each one-hour interval for the next day. In addition to hour-wise forecasts, day-wise PA duration, step, and calorie counts are also forecasted for every day of the next week. We infer the stillness and mobility of the smartphone user for the next day and week.