Analysis of Sensor Data Using Machine Learning Algorithms for Health Applications

Abstract

Data has become a fundamental part of every domain including healthcare. The continuous growth in technology and medical discoveries have resulted in an exponential increase in healthcare data. Additionally, the discoveries in the field of sensors and wearables have also contributed significantly as important data sources for healthcare data. The range of e-health sensors such as glucometers, accelerometers, gyroscopes, magnetometers, electrocardiograms (ECG), photoplethysmographs (PPG), electroencephalograms (EEG), etc. provides continuous data about the patient’s vitals, especially for chronic illness requiring active monitoring. This data obtained from sensors and wearables could be further utilized to gain actionable data insights for health workflow improvement using various machine learning (ML) algorithms.

Advancements in healthcare data analytics research have contributed to the development of various machine learning algorithms including neural network (NN) algorithms using e-health sensors/wearables data. These algorithms can detect/classify/ predict diseases such as type 1 and type 2 diabetes, arrhythmias, Parkinson’s, cardiovascular diseases, blood disorders, etc. In the same direction, this thesis has focused on addressing some of the technical challenges while analyzing the sensor data for disease prediction. At first, through a comprehensive review, the information on the wearable sensors that are used for chronic illness prediction is identified. It was identified that the sensors often used for chronic illness monitoring and are commercially available include accelerometers, gyroscopes, electrocardiograms, photoplethysmography, and blood glucose monitors. Furthermore, a multi-disciplinary review with a feasibility study was carried out to explore whether sensor data could be used to identify perceived loneliness levels.

For chronic diseases, such as diabetes, and Parkinson’s, that can affect all the vitals of the body, adding the data from multiple dimensionality or multiple sensors could help in getting comprehensive information on vitals and hence enhancements in prediction capability. To explore this possibility, ML analysis using multi-sensor data for predicting diabetes was carried out. The experimental analysis showed an increase of 4%-5% in accuracy using multi-sensor data. Furthermore, this thesis has also proposed a systematic methodology for combined hyperparameter evaluation required for complex ML, NN models, or those models involving lots of parameters/ features. The hyperparameter evaluation framework was evaluated for predicting Parkinson’s freezing of gait (FOG) symptoms.

Along with disease or its symptoms prediction, estimating the disease severity or symptom severity is also crucial for chronic disease. Severity analysis helps in identifying the stage of disease and in treatment planning as well. In this regard, this thesis has also investigated the possibility of predicting the symptom’s severity using sensor data. To evaluate this concept, two use cases were evaluated, one focusing on predicting the elderly’s perceived loneliness levels and another focusing on predicting the Parkinson’s FOG symptom severity. The ML analysis for perceived loneliness prediction showed that the social engagement and social behavior of a person could be highlighted using the mobility patterns with an accuracy ranging from 90%-98%. Similarly, for FOG severity estimation, the use of sensor data for labeling and predicting the severity levels achieved an accuracy of 88%. Additionally, the feasibility of home monitoring the FOG symptom severity has also been explored in this thesis.

With ML analysis, it becomes significant to also analyze the model’s computational efficiency, generality, scalability, and evaluate if the model is overfitting or underfitting. Moreover, for healthcare applications involving ML analysis, interpreting or explaining the model predictions is also important. This thesis has evaluated the multi-sensor ML models for their time complexity using performance curves between accuracy and convergence time, NN models for their generality or overfitting/ underfitting analysis using training curves and learning curves, and SHAP and PDP curves for interpreting models for feature and sensor relevance for severity estimation of FOG symptoms.

Other key observations made from this thesis were that the accelerometers and gyroscopes are the sensors that can be most likely used in many detection/prediction tasks including human activity recognition, chronic disease prediction, mobility or proximity analysis, etc. The accuracies for the prediction tasks will be improved if these sensors are combined with other sensor modalities. From the ML analysis carried out in the research studies, the ensemble algorithm namely XGBoost, was substantially more efficient in sensor data prediction.