A Novel thermal Imaging Based Transfer-Learning Model to Estimate Blood Pressure

Continuous monitoring of blood pressure (BP) is highly valuable for hospitalized and domestic patients; however, the existing non-invasive cuff-based BP monitoring system is discreet and applies artificial pressure on patient’ arms that makes the overall diagnostic process uncomfortable. The other invasive methods are interventional in nature, thereby the recuperating patient suffers from constant disturbances while measuring. Hence a cuff-less and contactless BP estimation system that allows caretakers to have easy access to measure BP of the patients at regular intervals without any disturbance is needed. A thermal imaging based continuous BP measurement system that is non-invasive, cuff-less, contact-less, non-disturbing, and targeted toward surgical, clinical, and domestic usage is proposed in this work. This system consists of a two-stage model wherein a pre-trained DenseNet-201 architecture is used to extract features from the thermal images and a 4 layered dense neural network (DNN) is employed on the extracted features to estimate the systolic blood pressure (SBP) and diastolic blood pressure (DBP) with high accuracy. For developing this system, a novel in-house data-set was established that consisted of thermal images of 20 healthy participants. The novel cuff-less blood pressure evaluation system presented a mean absolute error (MAE) of 2.27 mmHg and 2.51 mmHg for SBP and DBP respectively with similar standard-deviation (SD) metrics. The characterized error metrics of the proposed system are the lowest till date when compared to other prior work, and it also satisfied the BHS, AAMI, and IEEE standards with highest grade for further clinical and domestic usage.