Predicting Hypotension After Spinal Anesthesia Using Carotid Ultrasound and Clinical Variables

The induction of anesthesia is often followed by hypotension and the depth and duration of the hypotensive period influence the occurrence and severity of post-operative end-organ damage. The depth and duration of the hypotensive period might be reduced by predicting postinduction hypotension and administering fluids or medication at an earlier stage. While many methods to predict postinduction hypotension have been proposed, they require invasive measurements, lack explainability, or are not freely available. The recent emergence of carotid ultrasound patches makes carotid ultrasound a suitable alternative for patient monitoring, as it allows continuous non-invasive measurement of a central vessel. The present study aimed to build and validate a logistic regression model combining features from carotid ultrasound and clinical variables to predict postinduction hypotension. The model was trained and tested on a clinical dataset from adult patients scheduled for elective surgery under spinal anesthesia. At the preoperative holding area, carotid ultrasound imaging was performed and baseline vital signs were measured. Postinduction hypotension was defined as either 1) systolic blood pressure (SBP) reduction of >30% or mean arterial pressure (MAP) reduction of >20% from baseline, 2) an absolute SBP <90 mmHg or an absolute MAP <65 mmHg, or 3) the onset of hypotension-related symptoms. In total, 42 patients were included, of whom 19 (45%) became hypotensive after the induction of spinal anesthesia. The dataset was split in a ratio of 80:20% for training and testing. A total of 19 ultrasound and clinical features were studied and, after feature selection, five features were used to build the model. On the test dataset, the model had an AUROC of 0.88 and, using a threshold of 0.5, a sensitivity and a specificity of 75%. These preliminary results show that our logistic regression model might be able to predict postinduction hypotension. The next step is to test the proposed model on a large dataset and improve it if necessary. Moreover, other model types and features will be explored to evaluate if this improves the performance.