In Response

We appreciate the comments by Dr Bellini et al1 regarding our study.2 The advent of electronic health records and anesthesia information management systems (AIMS) has made highly granular perioperative data readily available, and it is an exciting period as we seek to understand the best ways in which these data can be utilized to improve patient outcomes. Machine learning methods provide new tools in our armamentarium to accomplish this task. However, the volume and complexity of these data present significant challenges that must be overcome for effective clinical tools to be developed from perioperative data. Bellini et al have noted an important consideration that some data are not collected automatically by the AIMS. For instance, fluid administration is manually recorded at sporadic intervals; in reality, fluids are administered continuously throughout the intraoperative period. Automated capture of these data may more accurately reflect the dynamics of intraoperative management and result in better models. Manually entered anesthetic management data are subject to errors, including inaccurate dosing and timing of medication data and omission of data.3 Another example of a modeling challenge relates to continuously recorded streams of physiologic data such as blood pressure. Decisions must be made regarding the optimal transformation of blood pressure values to data objects that can be entered into predictive models. Transformations such as minimum and maximum values, mean, median, and standard deviation have been utilized.4 Time-series approaches attempt to approximate the raw data stream and may provide utility for predicting perioperative outcomes.5 A caveat is that while complex data forms may improve model performance, interpretability may suffer and limit uptake in clinical practice. Efforts to develop effective clinical tools will require an interdisciplinary approach, including expertise in fields such as perioperative medicine, computer science and machine learning, bioinformatics, and biostatistics.6 Despite the numerous challenges, we agree and believe that there are tremendous opportunities ahead to realize the potential for big data and machine learning to improve perioperative management and patient outcomes. Minjae Kim, MD, MSDepartment of AnesthesiologyColumbia University Medical CenterNew York, New YorkDepartment of EpidemiologyMailman School of Public HealthColumbia UniversityNew York, New York[email protected] Xinyu Yan, MSDepartment of BiostatisticsUniversity of FloridaGainesville, Florida Guohua Li, MD, DrPHDepartment of AnesthesiologyColumbia University Medical CenterNew York, New YorkDepartment of EpidemiologyMailman School of Public HealthColumbia UniversityNew York, New York