Identification of Postoperative Pulmonary Complication Risk By Phenotyping Adult Surgical Patients Who Underwent General Anesthesia with Mechanical Ventilation

Abstract

Postoperative pulmonary complications (PPCs) remain a serious concern in health care. PPCs are associated with high rates of mortality and morbidity, intensive care admission, hospital length of stay, and resource utilization. PPCs are generally defined as any abnormality or condition involving the respiratory system that includes clinically significant dysfunction after surgery. PPCs are attributable to both modifiable and non-modifiable risk factors associated with characteristics of patients (e.g., age, sex, comorbidities), surgery (e.g., anatomical location, procedure length), and anesthesia (e.g., general anesthesia, mechanical ventilation). Although many PPC risks are not modifiable, intraoperative ventilation parameters (e.g., the fraction of inspired oxygen [FiO2], tidal volume [VT], sufficient positive end expiratory pressure [PEEP]) can be adjusted to reduce risk. Lung protective ventilation (LPV) has been adapted for intraoperative use to protect pulmonary parenchyma against ventilator-induced lung injury (VILI). LPV typically entails physiologic volume (i.e., lower VT) and pressure (i.e., PEEP), as well as optimal inspiratory time and alveolar recruitment maneuver. Despite growing evidence that intraoperative LPV can reduce the incidence of PPCs, questions remain regarding “how”, “when”, and for “whom” LPV can be used to reduce PPCs. Individualized care is one solution that could potentially minimize VILI leading to PPCs. Individualized care can be initiated by phenotyping patients based on observable nonmodifiable characteristics and modifiable characteristics as well as interactions among these characteristics. The goal of this dissertation was to advance knowledge around individualization of intraoperative ventilator parameters to reduce the incidence of PPCs. The first study (Chapter 3) answered “who” and “when” LPV in reducing the incidence of PPCs by leveraging the electronic health records (EHRs) and machine learning algorithms. We classified the adult surgical patients into phenotypes based on non-modifiable preoperative risks (e.g., age, sex, surgery type). First, a nonparametric machine learning algorithm, least absolute shrinkage and selection operator regression was used to select relevant variables. Then, a decision tree algorithm, classification and regression tree was then employed to identify phenotypic subgroups against the PPCs, and seven phenotypes were yielded for each outcome. This study suggested that phenotypes can be generated using the preoperative non-modifiable risks to predict PPCs. By extending the knowledge generated in chapter 3, the second study (Chapter 4) answered “how” to deliver individualized mechanical ventilation to optimize outcomes. We identified the optimal intraoperative mechanical ventilator parameters that were associated with the lowest incidence of PPCs for each phenotypic subgroup. The area under the receiver operating characteristic curve receiver operating characteristic curve was plotted using the estimates in the logistic regressions. The identified optimal values for VT and PEEP were associated with the lowest PPC incidence and the most desirable respiratory status postoperatively. What was considered optimal for VT and PEEP ranged between 5.11 and 9.31 ml/kg PBW and 5 and 11 cmH2O respectively. The results of this study suggested that intraoperative mechanical ventilator parameters should be adjusted based on a patient phenotype to optimize the postoperative outcomes. By utilizing machine learning algorithms and data-driven approach, this dissertation defined phenotypes based on non-modifiable preoperative PPC risk factors and identified the optimal ventilator parameters for each phenotypic subgroup. Simultaneously, we identified several future directions to advance our understanding on health care individualization to reduce incidence of PPCs and optimize PRS. This dissertation informs a future prospective study to vilify the optimal ventilator parameters, which may lead to the development of artificial intelligence to deliver individualized intraoperative mechanical ventilator parameters.