Perioperative risk prediction: Integrating clinical and methodological perspectives

Abstract

Elective surgery is widely considered safe, yet complications may still affect around one in six patients. This thesis explores how improved preoperative risk stratification can support better patient selection and facilitate earlier recognition and management of adverse outcomes. It brings together clinical studies on predicting postoperative infections, ICU requirements after esophageal surgery, and neurocognitive outcomes in older patients, alongside more fundamental methodological work on validating prediction models in the presence of missing data and second-order uncertainty in risk estimates. In part I, we examined perioperative risk is specific clinical contexts, focusing on overall complication risk in selected patient groups. In Chapter 2 and Chapter 3 we present the results of a systematic review of the literature aimed to assess the risk of perioperative thyroid storm, a rare but feared perioperative complication that is thought to occur more frequently in patients that are inadequately treated for hyperthyroidism before surgery. No studies were found that could reliably answer our research question or support the current guidelines to delay surgery until euthyroidism is achieved. In Chapter 4, we sought to determine the incidence of postoperative delirium and long-term psychopathology specifically in older patients, comparing these incidences with those in an age- and sex-matched cohort of older adults not undergoing surgery. Elderly surgical patients developed less depression 12 months following surgery compared to healthy controls. Postoperative delirium did not appear to influence the occurrence of long-term psychopathology. We then moved from these group-level summaries of risk to individualized predictions for postoperative infection and postoperative bed allocation using prediction modelling (part II). In Chapter 5 we performed a systematic review identifying 267 models for postoperative infection. Clinical usefulness of most models, however, remained questionable due to poor predictive performance, methodological shortcomings in their development, incomplete reporting and lack of validation. Through this systematic review, we were able to identify predictors that were frequently included in final models for postoperative infection. Building on these findings, Chapter 6 focused on the incremental value of red cell distribution width (RDW), a routinely available and inexpensive biomarker hypothesized to reflect chronic inflammation and oxidative stress, for improving risk stratification for postoperative infection in patients undergoing intermediate- to high-risk surgery. RDW did not improve prediction of postoperative infection risk. Chapters 7 and 8 focused on the development and external validation of a prediction model to distinguish patients requiring elective ICU admission following esophagectomy from patients that are eligible for PACU admission. Although the model had only moderate discriminatory performance, it was able to reduce ICU bed claims by at least 25%, compared to a default strategy of admitting all patients to ICU. Finally, in part III, we examined methodological aspects of robust risk modelling through two studies on validating prediction models in the presence of missing data and second-order uncertainty in risk estimates, and described the cohort profile of PLUTO, a perioperative biobank focused on prediction and early diagnosis of postoperative complications.