Sunny G Nijbroek1, Marcus J Schultz2,3,4, Sabrine N T Hemmes1,3. 1. Department of Anesthesiology. 2. Department of Intensive Care. 3. Laboratory of Experimental Intensive Care and Anesthesiology (L·E·I·C·A), Amsterdam University Medical Centers, University of Amsterdam, Amsterdam, the Netherlands. 4. Mahidol Oxford Tropical Medicine Research Unit (MORU), Mahidol University, Bangkok, Thailand.
Abstract
PURPOSE OF REVIEW: Prediction of postoperative pulmonary complications (PPCs) enables individually applied preventive measures and maybe even early treatment if a PPC eventually starts to develop. The purpose of this review is to describe crucial steps in the development and validation of prediction models, examine these steps in the current literature and describe what the future holds for PPC prediction. RECENT FINDINGS: A systematic search of the medical literature identified 21 articles reporting on prediction models for PPCs. The studies were heterogeneous with regard to design, derivation cohort and whether or not a validation cohort was used. Furthermore, as definitions for PPCs varied substantially, PPC rates were quite different. One-third of the studies had a sufficient sample size for building a prediction model. In most articles, an internal validation step was reported, suggesting a good fit. In the four articles that reported an externally validation step, in three the prognostic model performed less well in external validation. The ARISCAT risk score was the only score that kept sufficient predictive power in external validation, albeit that the sample sizes of the cohorts used may have been too small. Analysis by machine learning could help building new prediction models, as unbiased cluster analyses could uncover clusters of patients with specific underlying pathophysiological mechanisms. Adding biomarkers to the model could optimize identification of biological phenotypes of risk groups. SUMMARY: Many predictive models for PPCs have been reported on. Development of more robust PPC prediction models could be supported by machine learning.
PURPOSE OF REVIEW: Prediction of postoperative pulmonary complications (PPCs) enables individually applied preventive measures and maybe even early treatment if a PPC eventually starts to develop. The purpose of this review is to describe crucial steps in the development and validation of prediction models, examine these steps in the current literature and describe what the future holds for PPC prediction. RECENT FINDINGS: A systematic search of the medical literature identified 21 articles reporting on prediction models for PPCs. The studies were heterogeneous with regard to design, derivation cohort and whether or not a validation cohort was used. Furthermore, as definitions for PPCs varied substantially, PPC rates were quite different. One-third of the studies had a sufficient sample size for building a prediction model. In most articles, an internal validation step was reported, suggesting a good fit. In the four articles that reported an externally validation step, in three the prognostic model performed less well in external validation. The ARISCAT risk score was the only score that kept sufficient predictive power in external validation, albeit that the sample sizes of the cohorts used may have been too small. Analysis by machine learning could help building new prediction models, as unbiased cluster analyses could uncover clusters of patients with specific underlying pathophysiological mechanisms. Adding biomarkers to the model could optimize identification of biological phenotypes of risk groups. SUMMARY: Many predictive models for PPCs have been reported on. Development of more robust PPC prediction models could be supported by machine learning.
Authors: Liselotte Hol; Sunny G L H Nijbroek; Ary Serpa Neto; Sabrine N T Hemmes; Goran Hedenstierna; Michael Hiesmayr; Markus W Hollmann; Gary H Mills; Marcos F Vidal Melo; Christian Putensen; Werner Schmid; Paolo Severgnini; Hermann Wrigge; Marcelo Gama de Abreu; Paolo Pelosi; Marcus J Schultz Journal: BMC Anesthesiol Date: 2022-01-07 Impact factor: 2.217