Realtime Vital Sign Monitoring Integrated with Neonatal Disease Prediction System for Neonatal Intensive Care Unit

Abstract

Neonates or infants under 28 days, are vulnerable to many diseases and complications that may cause death. This is because their immune system is very low until they reach three months or above. Child mortality is very high in developing countries compared to those economically developed countries due to limitation in neonatal intensive care unit (NICU) that include lack of continuous follow up, lack of medical expertise and lack of the implementation of advanced technologies that have a major role in reducing child mortality. The main objective of this study is to design an integrated internet of thing (IOT) system with machine learning approach for a continues real-time vital sign monitoring and early prediction of neonatal disease in NICU that include early onset sepsis, respiratory distress syndrome, jaundice and hypothermia. To achieve the aim of the study a small portable vital sign module was designed using temperature, heart rate, blood oxygen level and respiration sensors. Wemos D1 mini was used to acquire and process sensor data as well as to display the results on liquid crystal display. Also, Wemos D1 mini parallelly send sensor data to a database server for storing and providing a real-time continuous monitoring of temperature, heart rate, blood oxygen level and respiration rate. A real time vital sign data was obtained by developing an IOT network that connect the designed module with the database system. In addition, to develop a prediction model a clinical data composed of symptom list, vital sign, history and physical examination were collected from neonatal intensive care unit. The raw data collected underwent a simple imputation technique to handle missing values. To balance the total data per class, Synthetic Minority Oversampling Technique was used. In addition, one-hot and label encoding were utilized to convert categorical data into numerical form. The pre-processed data was used to train and test different models including sequential model, random forest, gradient boosting, extreme gradient boosting, decision tree, support vector machine and logistic regression. The two systems were integrated using a custom-designed graphical user interface. The prototype was tested using 20 healthy adults, and according to bland Altman statistical test the mean bias and limit of agreement 95% confidence interval were: temperature: -0.025˚C ± 0.445 ˚C, heart rate: -0.3 ± 2.55 beat per minute, oxygen saturation: -0.15 ± 1.46 %, respiration rate: 0.25± 3.04 breath per minute. For early disease prediction, best results were obtained using iv | P a g e random forest model with accuracy, precision, recall and f1-score of 96%, 96%, 97% and 96%, respectively. Our experimental results show that the system provide accurate results in predicting NICU diseases that include early onset sepsis, respiratory distress syndrome, jaundice and hypothermia. This technology has the potential to reduce child mortality, particularly in developing countries with limited access to medical expertise and advanced technologies. Further research is needed to implement this system in real-world clinical settings.