Predicting Neuromuscular Recovery in Surgical Patients Using Machine Learning

Overview

Despite emerging efforts to decrease residual paralysis and postoperative complications with the use of quantitative neuromuscular monitoring and reversal agents their incidences remain high. In an optimal setting, neuromuscular blocking agents are dosed in a way that there is no residual block at the end of surgery. The effect of neuromuscular blocking agents, however, is highly variable and is not only influenced by their dose, but also by several patient-related factors such as muscle status, metabolic activity, and anesthesia management. Accordingly, the duration of action is difficult to predict. The PINES project will use artificial intelligence methods to develop a model that can accurately predict the course of action of neuromuscular blocking agents. It will be used to predict time to complete neuromuscular recovery (train-of-four \[TOF\] ratio \>0.9) and may provide as a decision support in the individual management of timing and dosing of neuromuscular blocking drugs and their reversal agents. In a secondary analysis, the association between the choice of neuromuscular blocking agent and postoperative pulmonary complications will be evaluated.

The objective of the PINES project is to identify a model that can accurately predict 1) time to complete neuromuscular recovery, 2) optimal timing and dose of neuromuscular blocking agents at each time point during surgery, and 3) TOF ratio at the estimated end of surgery to assess residual paralysis. Furthermore, a prospective clinical pilot study will be conducted to compare anesthesiologist-predicted neuromuscular recovery with that of the algorithm. The project consists of two main objectives: I. Big data analysis * Establishing a data warehouse: Electronic registry data will be used. * Generation of prediction models: Classification models will first be used to identify and weight the relevant parameters collected during premedication and intraoperatively. These will form the basis for the training cohort, which can then be used to carry out a simulated real-time analysis of the data. To compare the models, the loss functions mean squared error, mean absolute error and Huber Loss will be calculated. II. Prospective comparison of the prediction: machine-learning model vs. anesthesiologist Using the validated final prediction model with the best accuracy, we will perform a prospective clinical pilot study. The cohort will include prospectively enrolled adult surgical patients undergoing general anesthesia with a single dose of rocuronium for neuromuscular blockade. For each enrolled case, both the PINES algorithm and an experienced anesthesiologist will estimate the time to neuromuscular recovery, defined as a train-of-four (TOF) ratio \> 0.9. At anesthesia induction, following administration of the neuromuscular blocking agent, participating specialist-level anesthesiologists will prospectively estimate the time in minutes until recovery of neuromuscular transmission. The PINES machine-learning model will generate its prediction. The actual recovery time will be determined from the continuously recorded intraoperative TOF measurements. The agreement between the predicted and observed recovery times will be assessed by calculating the difference between predicted and actual values, as well as by determining inter-rater correlation coefficients comparing anesthesiologist predictions, algorithm predictions, and the measured recovery times. In a secondary analysis, there will be evaluated whether the choice of neuromuscular blocking agent influences postoperative pulmonary complication risk in adult patients. Confounding will be addressed using statistical methods based on a causal inference framework.