Integration of a wearable IoT-based smart healthcare system with predictive modelling.

Abstract

Abstract: The usage of the Internet of things (IoT) equipped with intelligent machine learning (ML) or fuzzy logic (FL) systems in the medical sector is becoming a widespread normality. This following the implications of the 4th and 5th industrial revolutions (5IRs) aiming to integrate technology with the way we live and work. Many research areas have focused on incorporating individual elements of an intelligent IoT based system. However, none have holistically looked at encompassing a multi-faceted sensory device with an IoT framework backed by both ML and FL capabilities for unique monitoring and diagnostics. This implies that there is an opportunity to create such a product targeted for the medical sector. The developed system of systems (SoS) discussed therefore aims at eliminating the barrier to patient care outside of a hospital setting through the use of real time patient diagnostics. The study was approached by first establishing a streamlined IoT infrastructure. A multi-faceted Arduino based sensory device was then developed to collect user parameters. The electronics were housed in a common of the shelf (COTS) casing satisfying the criteria of wearability. Bluetooth connectivity then allowed for the transmission of sensory data from the WBAN to the IoT smartphone gateway. Sensory parameters being measured were electrocardiography (ECG), electromyography (EMG), body temperature (BT), infrared (IR) output, pulse rate (PR) as well as environmental humidity readings. Blood sugar (BS) levels were obtained non-invasively by calibrating the IR output voltage with that of a calibrated glucometer to obtain a straight line equation representing the relationship between the sensor ADC output and glucose in the blood. A Java based mobile application (MA) was then developed, which allowed for the processing of the sensory data before storage in a local SQLite database. A two interface MA for use by registered doctors and patients was necessary to allow for data sharing and security. Registration and login for the MA was done through a NoSQL Firebase database. A JDBC was established to enable the transmission of data to a MS Azure Structured Query Language (SQL) database. This served as the cloud interface within the IoT network layer. The MA is able to integrate with a phone’s Global Positioning System (GPS) to allow for simultaneous tracking of patients. A return application programming interface (API) connection between MS Azure and the developed application layers was then created. Here the mobile doctor interface as well as a developed node.js based web application (WA) served as the mediums through which health practitioners could access patient data. ML models were developed using the MS Azure ML Classic Studio suite which allowed for a real time analysis of received data through deployment and subsequent consumption of data using POST Requests via the MA and a Java software development kit (SDK). An accuracy of 92% was achieved for the stroke prediction model based on the boosted decision tree algorithm. Various ML models were analyzed to ensure that a high precision was obtained while preventing overfitting. Additional FL models were also developed, which took into consideration unique sets of vitals combinations to create a rule base depicting the patient health status and health risk. v This data was compared to the intuition of a doctor and received an 87% accuracy for model 1 which took into consideration a patient’s PR, BS, BT and age to predict the health status of a patient. The model was then compared to the Modified Early Warning Score (MEWS) rating, a popular measure of health risk utilized in the medical field. The results of comparison also yielded an accuracy of 87%. A second FL model was then developed looking at the effect of environmental conditions on the risk rating of patients. Here input variables of BT, age and humidity readings were used to determine their effect the on the risk rating of a patient. This model scored an 80% accuracy when compared to the expertise of a physician. Both models were programmed onto the MA to predict the patient’s health status. Both models were also re-developed on MATLAB software to simulate the effect of various input variables on the response variable. Overall the designed system was able to possess around 15 of the typical features found in smart wearable systems which far exceeded the features of those devices it was compared to. The designed system satisfied the requirement of a feature rich experience while also satisfying the criteria of cost effectiveness.