Time series forecasting of cardiopulmonary signals during exercise

Abstract

quotExercise-based evaluation of the cardiopulmonary system is always advisable over other newlineinvasive and costly diagnostics procedures. It has enough potential to judge the fitness of newlinehealthy persons as well as to detect symptoms of abnormality earlier. In the incremental newlineexercise, amount of workload increases at every few minutes interval, so after initial newlinestages, the amount of workload to the cardiac system is above average so sometimes it may newlineoverload the cardiopulmonary system and this is not tolerable every time especially when newlinethe subject is a cardiac patient. This research work presents a new approach for forecasting newlineof cardiopulmonary signals after the premature end of the test. Here, three models are newlineimplemented which are the Adaptive filter, the Autoregressive moving average with newlineexogenous terms (ARMAX) and the nonlinear ARX (NLARX), for time series forecasting newlineof signals like Instant heart rate (HR) and respiration rate (RR). The models are newlineimplemented such that it utilizes subject s own past and current responses to forecast newlinefuture response, there is no need of another database for training or any other purpose. newlinePerformance of these models are tested on normal as well as abnormal cardiac subjects. newlineThe normal database is collected with the help of young subjects facing physical stress newlinelevel of exercise protocol BRUCE on Treadmill machine and abnormal cardiac patients newlinedatabase are collected from the online physiological database of PhysioNet. After newlinevalidation of results and statistical analysis of forecasting errors, the NLARX model is newlinefound to be more accurate and reliable for forecasting of cardiopulmonary signals like newlineInstant HR and Instant RR for both kinds of subjects, Normal as well as Abnormal. The newlineperformances of these three models ranked in descending order are: NLARX, ARMAX, newlineand Adaptive filter. Additionally, the prediction algorithm is implemented for single step newlineand multi-step-ahead (five step ahead) prediction of Instant RR with the help of artificial newlineneural network structure- nonlinear autoreg