Predictive analytics in health monitoring

Thumbnail Image



Journal Title

Journal ISSN

Volume Title


University of New Brunswick


Predictive analytics in healthcare can prevent patients emergency health conditions and reduce costs in the long term. Accurate and timely anomaly predictions focusing on recent events can save lives. Nevertheless, for such accurate predictions, machine learning algorithms require processing long-term historical big data, which is infeasible in wearable devices due to their memory constraints and low computing power. Current techniques either ignore a large amount of historical data or convert temporal sequences to pattern sequences, eliminating valuable properties for prediction such as time and recency. In addition, missing values in data collection can impair the prediction. Hence, the motivation of this research is to efficiently model historical data with missing values in a precise form of multivariate temporal sequences to detect and forecast emergency events. The proposed model is named as life model (LM). LM creates a new concise sequence to represent the history and the future as an intensity temporal sequence (ITS) tensor. LM maps arbitrary-length multivariate discrete timeseries data to another concise sequence, called multivariate interval sequence (MIS). ITS and MIS retain the original data properties such as time, recency, and scale, without being much susceptible to missing values. Since long short-term memory (LSTM) recurrent neural networks are proved to be effective models for modeling sequence data, the LM algorithms and their properties enable ITS and MIS tensors to train LSTM and other machine learning techniques efficiently in order to predict in real-time, even in the absence of some values. LM is tested to predict and forecast emergency event such as the mortality of a patient from the MIMIC III intensive care unit dataset. Based on their diagnosis and procedure codes over a span of 11 years, the model achieved 84.2% and 99.6% accuracy on 34k and 10k patient records respectively. In addition, the LM model is tested to predict the approximate time of certain human activities, with different granularity of seconds and up even to years. When tested on the URFD fall dataset, the experimental results show that, compared to a previous study using a complex LSTM network, LM achieves the same 100% accuracy in fall prediction using 80x less weight parameters and computing power. LM is observed to forecast human fall up to 14 seconds in advance with 86.96% accuracy with all available data and 85.56% accuracy with 50% missing values. Finally, a new LM-powered predictive health analytics and real-time monitoring schema (PHARMS) is developed which uses deep learning for predictive analysis in a medical internet of things environment using wearable devices.