PrEventing PostoPERative Pulmonary Complications by Establishing a MachINe-learning assisTed Approach

Overview

Postoperative pulmonary complications (POPC) are common after general anaesthesia and are a major cause of increased morbidity and mortality in surgical patients. However, prevention and treatment methods for POPC that are considered effective, tie up human and technical resources. The aim of the planned research project is therefore to enable reliable identification of high-risk patients on the basis of a tailored machine learning algorithm using perioperative clinical routine data and sonographic imaging data collected in the recovery room. The randomized clinical trial will include 512 patients undergoing elective surgery in general anaesthesia. The primary outcome will be the development of POPC. The goal of the study is to detect postoperative pulmonary complications before they become clinically manifest.

The incidence of postoperative pulmonary complications (POPC) is reported to be 9-40%, depending on the surgical procedure. Various preoperative risk factors are known, but usually cannot be modified. A major problem of older publications was that for a long time there existed no clear definition of the outcome parameter "pulmonary complication". It was not until 2018 that a standardised definition was developed by the Standardised Endpoints for Perioperative Medicine (StEP) collaboration. Due to the high clinical relevance - POPC are the main cause of postoperative morbidity and mortality - clinical scoring systems for the preoperative prediction of POPC have been developed, but their predictive quality still needs to be improved. The currently best evaluated score for predicting postoperative pulmonary complications (ARISCAT: Assess Respiratory Risk in Surgical Patients in Catalonia) has sufficient sensitivity but lacks specificity. Therefore, machine learning methods for determining risk from preoperative routine data are also being tested. Sonography is becoming increasingly important as a non-invasive examination method that can be performed at the bedside. Various sonographic scores and models have already been developed to predict pulmonary complications. Image processing methods and machine learning, in particular deep learning are also increasingly being used in ultrasound diagnostics. A combination of routine clinical data and imaging data to develop a machine learning algorithm has not yet been tested. However, augmented algorithms using pre- and intraoperative clinical information in addition to ultrasound imaging promise better predictive accuracy than the respective individual methods. In addition, prospective clinical evaluation of machine learning algorithm-based prediction models is lacking to date, although they show good values for "area under the receiver operating characteristic" (AUROC), accuracy and precision in the respective test and validation datasets, which are considered common measures of the predictive quality of such models. Measures for the prevention of POPC are known, but are probably not consistently applied in clinical routine due to the increased demand, especially for human resources. Therefore, the aim of the study is to identify patients at risk of POPC on the basis of a machine learning algorithm. All patients are undergoing the same study protocol to develop the machine learning model. Perioperative clinical routine data are going to be assessed as per standard. Postoperatively, a standardized lung sonography is going to be performed in the recovery room. Patients will then be visited on the ward on postoperative day 1, 3 and 7 for clinical examination to detect POPC according to the criteria elaborated by the StEP- collaboration. According to the case number calculation, 512 adult patients undergoing elective, surgical procedures under general anaesthesia are going to be included. Perioperative routine data will be assessed and stored in a hospital-internal database, as well as data from postoperative clinical examination. Image data from lung sonography will be archived in the PACS for further processing. Based on the collected data, a machine learning algorithm based on neural networks will be trained to predict POPC. The model is created with the anonymized data using the statistics-oriented programming language R and the framework TensorFlow, a deep learning software library based on the programming language Python. The prediction quality of the created prediction model is assessed using the area under the receiver operator characteristics (AUROC) as well as the area under the precision recall curve (AUPRC) and compared with the values of the ARISCAT score, a common score to estimate the risk of POPC. Precise risk assessment by means of an augmented machine-learning algorithm that uses clinical routine as well as imaging data has great potential to improve patient outcomes and could also help to reduce health care costs.