Noninvasive Identification of Hypotension Using Convolutional-Deconvolutional Networks

High-frequency identification of a patient's hypotensive state allows for early notification of adverse events and long-term trends. Risks of complications such as heart attack, acute kidney injury, and mortality increase with duration of hypotension during surgery, and a hypotensive state can affect the appropriate medication choices and dosages for congestive heart failure patients. Current methods for identifying hypotension are based on blood pressure cuff measurements, which are low-frequency and must be manually collected, or catheterized blood pressure sensors, which are invasive, painful, and not necessarily usable for the youngest and smallest neonatal patients. This paper explores the potential of replacing the high-frequency hypotensive state produced by the invasive arterial catheter with a high-frequency projection from a fusion of multiple noninvasive sensors. These noninvasive sensors are available for a large majority of hospital patients, and have a lower risk of adverse effects ranging from patient discomfort to site infection. In addition, using multiple sensor inputs and a robust model allows for higher-reliability identification than single-sensor architectures. Our results demonstrate that by using a single flexible convolutional-deconvolutional neural network architecture, a patient's hypotensive state may be reconstructed from any combination of the input sensor channels, with fidelity increasing in the number of available inputs.