Predicting Systolic Blood Pressure Using Machine Learning

Abstract

In this paper, a new study based on machine learning technique, specifically artificial neural network, is investigated to predict the systolic blood pressure by correlated variables (BMI, age, exercise, alcohol, smoke level etc.). The raw data are split into two parts, 80% for training the machine and the remaining 20% for testing the performance. Two neural network algorithms, back-propagation neural network and radial basis function network, are used to construct and validate the prediction system. Based on a database with 498 people, the probabilities of the absolute difference between the measured and predicted value of systolic blood pressure under 10mm Hg are 51.9% for men and 52.5% for women using the backpropagation neural network. With the same input variables and network status, the corresponding results based on the radial basis function network are 51.8% and 49.9% for men and women respectively. This novel method of predicting systolic blood pressure contributes to giving early warnings to young and middle-aged people who may not take regular blood pressure measurements. Also, as it is known an isolated blood pressure measurement is sometimes not very accurate due to the daily fluctuation, our predictor can provide another reference value to the medical staff. Our experimental result shows that artificial neural networks are suitable for modeling and predicting systolic blood pressure.

Improving Accuracy of Patient Synthetic Data for Testing Medical Cyber-Physical Systems

“…The information in this database refers to patients admitted to Intensive Care Units (ICU) whose data were collected from bedside monitors and Vital sign Predictor variables BP [6] Environment temperature, age, gender, body mass index (BMI), alcohol consumption, smoking, cholesterol and blood glucose. BP [5] Alcohol consumption, age and exercises. BP and HR [3] Weight, age, blood pressure at rest, heart rate at rest, exercise intensity, type of exercise and oxygen consumption.…”

“…Finally, for the body temperature (bt value) variable, we used the Gama regression model, whose canonical link function was presented previously in (5). After adjusting the regression model for BT (M LG BT ), we obtained the R 2 * = 0.755.…”

“…The regression model chosen for the heart rate variable (hr value) was the Gama Linear Model with canonical link function given by (5). This regression model also belongs to the MLGs class and, when related to heart rate, is what best represents the variability of the data found in the sample.…”

“…For the systolic blood pressure (sbp value) variable, the Gama regression model was also used, whose canonical link function was previously shown in (5). Other models have been tested, however, this was best represented the variability of the data.…”

“…Van Heusden et al [4] proposed an artificial pancreas model for patients with type 1 diabetes mellitus, with the goal of improving glucose control in such patients. Wu et al [5] and Bhaduri et al [6] used artificial neural networks to investigate the correlation between blood pressure and some variables such as alcohol consumption, body mass index (BMI), age, and exercise, thereby building patient models to represent specific behaviors of the human body. Finally, Khan et al [7] provided a glucose control system to be used to prevent hypoglycemic episodes, in which the patient model (i.e., the artificial pancreas) establishes a relationship between heart rate and blood glucose level.…”