Letter: Big Data Research in Neurosurgery: A Critical Look at This Popular New Study Design

Abstract

To the Editor: We were interested to read the recent article by Oravec et al1 about the increasing use of administrative databases in neurosurgical research. The authors surveyed the neurosurgical literature for articles that use administrative databases for the purpose of answering questions primarily regarding patient outcomes and healthcare costs. The authors highlight some important limitations of the currently available databases, including lack of neurosurgeon involvement in their design and coding inconsistencies. Nevertheless, there are some points in the article that merit discussion. Oravec et al1 use the term “Big Data” to describe the use of administrative databases, such as the National Inpatient Sample or the National Surgical Quality Improvement Program database. While one of the merits of database use is indeed sample size, volume is only one component of what defines “Big Data.” The 3 other commonly cited characteristics of Big Data are variety, velocity, and veracity.2 Variety refers to the different types of information that are, in a health care setting, associated with each patient. Some examples include imaging, pathology, vital sign monitoring and electronic health record (EHR), among many others. Velocity refers to the rate of data accumulation and analysis; a perfect example of high rate of data accumulation in neurosurgery is the data generated from continuous intracranial pressure (ICP) monitoring or electroencephalogram (EEG) recordings. Lastly, veracity refers to quality of the data and is closely related to the concept of internal validity. All in all, the term “Big Data” more so describes the approach to data science defined by these 4 characteristics rather than the actual sample size (ie, number of patients represented). Particularly since the widespread adoption of EHR, the use of Big Data computational approaches has been of much interest in the medical community.3-6 Furthermore, the nature of Big Data research lends itself well to computational analysis using machine learning and deep learning methodology,6 which eliminates the problem of “P-hacking” as highlighted by the authors and warned against by the American Statistical Association.7 Neurosurgical pathologies lead to a range of clinical manifestations and patients are monitored by a variety of different techniques allowing for the integration of large set of multimodal data. This variety and volume of data make the field of neurosurgery particularly well suited to benefit from the appropriate use of Big Data methodology such as deep learning. One example is the application of machine learning to predict elevations in ICP using continuous data streams from continuous monitoring of ICP and vital signs.8,9 Additionally, some have applied these methods to EEG readings to automate seizure detection and localize epileptogenic focus.10,11 In the field of neuro-oncology, machine and deep-learning approaches have been used to analyze magnetic resonance images to classify and grade tumors, and predict survival.12-14 These are only some of the applications of Big Data methods that are currently under investigation in neurosurgery, and specifically by our group at the Computational Neuroscience Outcomes Center, and a more in depth review can be found elsewhere.15 Though these applications show great promise, they face challenges that limit prompt clinical integration. While healthcare certainly has access to significant data storage and computational power to allows for handling volume and velocity, these are not often available at the bed-side that may limit the real-time use of machine learning algorithms for prediction of ICP changes, for instance. Additionally, algorithms that are developed using training data sets from one institution may not perform with the same accuracy on data from a different institution. This can be attributable to differences in data acquisition techniques, data processing, and even in the range of underlying pathologies seen in different institutions. This relates to an issue raised by Oravec et al1 regarding the reproducibility of computational data analysis. We advocate that an open-source sharing of computational methods and deidentified heterogeneous datasets are important ways to address this issue. Frameworks have been developed within the Food and Drug Administration to regulate algorithms trained on Big Data sources and meant for use in the clinical setting under the umbrella of Software as a Medical Device. With regards to administrative databases, Oravec et al1 rightly bring up concerns about the quality of diagnosis coding that may impact the internal validity of studies using these data. Quality improvement efforts by the institutions that manage these databases, including the involvement of neurosurgeon input in database design, are one approach to address this issue. An open-source designation of neurosurgical diagnosis codes and relevant complications can also help standardize the way the field uses these datasets. Additionally, like other types of retrospective data, analysis drawn from administrative databases can be used for descriptive and inferential analysis, but cannot establish causality. This should be appropriately mentioned as a limitation in any paper that uses this type of data. These limitations, however, do not obviate the utility of administrative databases in answering particular types of questions and as hypothesis-generating tools. Descriptive and inferential analysis aside, the volume provided by administrative databases, which is unparalleled by any prospective dataset, is especially useful for the generation of predictive models. Ultimately, our goal is to highlight the fact that administrative databases are an example of only one type of data that falls within the umbrella of “Big Data,” and that we have only begun to scratch the surface of what the data science, powered by Big Data and artificial intelligence, can offer the neurosurgical community and our patients. Disclosure The authors have no personal, financial, or institutional interest in any of the drugs, materials, or devices described in this article.

Ähnliche Arbeiten