Human activity recognition identifies different human body movements and estimates the dwell time for each activity which can be widely used for behavior analysis and rehabilitation assessment in the areas of elderly care, health care, assisted living, and athlete training. We perform preliminary investigation of recognizing human activities (including real-time fall detection and its direction) using wearable technologies and machine learning algorithms. From daily activities’ point of view, the most dangerous situation for the post-stroke patients or elderlies is fall. To detect fall and assess the severity of fall as early as possible can enable timely intervention.
In our research, a human subject wears an IMU sensor on the chest to acquire raw accelerometer and gyroscope data for each activity. Five different activities, such as sitting/standing, lying down, going upstairs, going downstairs, walking, and falling (in 4 directions) are considered. After acquiring raw motion data, feature construction is executed. Fifty four statistical features are generated based on different statistics, such as Range, Mean, Absolute Mean, Mean Cross Rate, Zero Cross Rate, Standard Deviation, Covariance, and Mean Trend. Feature selection is performed afterwards to find a subset of the original full feature set. Three different filter-based feature selection methods are used here, i.e., kBest feature selection, Relief-F feature selection, and robust feature selection. Finally, activity classification is on duty to recognition different human activities based on the input features built for each activity.
We used data analytics techniques to build a predictive model to forecast hospital discharge status of stroke patients in the State of Tennessee. Tennessee Department of Health provided hospital discharge data corresponding to stroke patients. In other words, we want to predict whether a stroke patient should be discharged after hospitalization to home or not home. In this way, we can perform personalized health care recommendation, improve quality of health care, and reduce cost of health care. We consider the following potential predictors, i.e., sex, age, race, ethnicity, stroke type, comorbidities (diabetes, heart disease, hypertension, peripheral arterial disease, chronic kidney disease, hyperlipidemia, stroke, arrhythmia, and depression), bill type, admission type, admission source, and primary and secondary payer classes. Three different machine learning methods, i.e., logistic regression, random forest, and support vector machine, are exploited to build predictive models. The following figures show receiver operating characteristic (ROC) curves of predictive models for derivation set and validation set, respectively.
ROC curve for derivation set. ROC curve for validation set.