Empirical pattern driven dengue fever severity prediction model

Abstract

Most dengue fever infection is asymptotic or pre-symptomatic disease and the severity develop over time. Dengue patient managed under Group A (outpatient) required for home monitoring. During home monitoring, the patients might not know whether the vital signs measured are correct or good enough and the result might not be convincing or can be misinterpreted. For example, in dengue, as the fever goes away, the patients might think they already recover but actually they might enter the critical phase when most mortality occur. Thus, an early detection and consistent monitoring of dengue fever infection at early stage is demanding. Hemodynamic or vital signs assessment such as body temperature, heart rate and blood pressure measurement are the usual assessment done by doctor when they are examined. These parameters will help doctors assess the condition of the patient. In dengue, these physiological parameters together with the blood profile are used as the reference parameters for dengue progression and dengue treatment. Thus, the main objective of this research is to develop a dengue severity prediction model based on vital signs and blood profile. The other objectives in this research are; to acquire the retrospective dengue patients in the clinical setting, to pre-process and analyze the data using statistical method, to model the extracted features using machine learning techniques, to established the predictive modelling using the best classifier and lastly to design a vital signs monitoring system that able to collect near real time vital sign data in clinical setting or home setting. In this research the dengue patients’ data are pre-processed using SMOTE technique and analyzed using IBM SPSS Statistics 21 and the extracted features would model using machine learning techniques in MATLAB R2020b. The statistical analysis used the Kruskal-Wallis and ANOVA for parametric and non-parametric data and Bonferroni-Dunn and t-test as the post hoc analysis. The vital signs monitoring system is able to collect near real time vital sign data, developed using e-health platform sensor, Arduino Uno and LabVIEW data dashboard internet of thing. The result also is automatically saved in Microsoft Excel format for future reference. As the result, 62 dengue patients’ data, with 8 parameters from three dengue categories were collected. All these data were recorded from the first day they visited emergency department until they were discharged. White blood cell was found as the variable that could be used to differentiate two dengue categories; severe dengue versus dengue fever and severe dengue versus dengue fever with warning signs, whereas systolic blood pressure and diastolic blood pressure were the variables that could be used to differentiate dengue fever and dengue fever with warning signs from day 3 to day 8 of the illness. The support vector machine was found as the best classifier for this prediction model. Day 3 of illness predicted with 89.7% accuracy, day 4 of illness predicted with 89.7% accuracy, day 5 of illness predicted with 82.1% accuracy, day 6 of illness predicted with 92.3% accuracy, day 7 of illness predicted with 87.2% accuracy and day 8 of illness predicted with 84.6% accuracy. These prediction model could assist the physicians in identifying the dengue severity category and provide appropriate treatment, which could reduce the complexity in dengue management.