ThermalBP: A Unified Thermal Imaging Model to Estimate Blood Pressure

This paper proposes a generalized non-invasive machine-learning (ML) based system to monitor blood pressure (BP) with the eventual aim of replacing the traditional cuff-based systems. The existing cuff-based designs are interventional and apply pressure on the hands of the patient which makes the diagnosis uncomfortable. The previous attempts to design a non-invasive diagnostic procedure have been performed under controlled environmental conditions, which is not feasible in a real-world scenario. The proposed design aims to generalize the environment in which the experiment is conducted. The proposed model incorporates the subject's data in the form of additional features including height, weight, and age of the individual under test, and ambient temperature, which in hindsight becomes crucial in improving the prediction accuracy of the BP. The system consists of a two-staged transfer learning-based model, where the features are initially extracted from a pre-trained DenseNet161 architecture and fed to a 5-layered Dense Neural Network along with the additional features in the second stage. The proposed cuff-less design presented a mean absolute Error (MAE) of 1.79 mmHg for SBP and 1.5 mmHg with their respective standard deviation (SD) metrics. The characterized error metrics are the lowest to date when compared to previous attempts. The novel in-house dataset and model is made freely available for further usage to the researchers and developers community.