Kayleigh Kenny (Philadelphia, PA / US), Victoria Gruen (Philadelphia, PA / US), Meghan Willner (Philadelphia, PA / US), Craig Maddux (Philadelphia, PA / US), Ethan Jacob Moyer (Philadelphia, PA / US), Brandon Foreman (Cincinnati, OH / US), Eric Rosenthal (Boston, OH / US), Professor DaiWai Olson (Dallas, TX / US), Dick Moberg (Philadelphia, PA / US)
Abstract-Text (inkl. Referenzen und Bildunterschriften)
Background Common Data Elements (CDEs) were developed to harmonize data sets in clinical trials. Although they are successful in harmonizing multi-study data, they fall short in providing relevant annotations for machine learning algorithms in research applications and routine clinical care. Our focus in this study was to identify the gaps in harmonization of multimodal physiological data from neurocritical care (NCC) patients.
Methods
With prior work in data harmonization, nomenclature standards, and a small group of stakeholders, we performed a gap analysis between CDEs and terms needed to characterize data acquisition and data analysis. We reviewed National Institute of Neurological Disorders and Stroke (NINDS) CDE terms for traumatic brain injury from 18 modalities across nine measurement categories. After tabulating the CDEs, we separated them based on if they were related to data acquisition (e.g. intracranial pressure device type, C01572) or data analysis (e.g. intracranial pressure mean daily measurement, C01575). We cross-referenced each CDE set with monitoring scenarios and existing standard care guidelines to determine additional terms needed to correctly characterize the acquisition and analysis of the data. We then performed a second gap analysis with the Brain Trauma Foundation updated CDEs for physiology and big data with 21 modalities across the same nine measurement categories.
Results
After conducting our gap analysis, we observed that 21 modalities lacked critically important terms with respect to CDEs to describe data acquisition and analysis. We found that for 20 modalities, additional data acquisition terms were necessary to improve standardization of data across institutions. We also identified and added terms to 20 modalities for data analysis. EEG was the only modality considered adequately contextualized in the original CDE database. Other analysis terms were omitted for being too study specific. For example, "intracranial pressure (ICP) episode greater 20 mmHg count" (CDE C01577) was removed as standards of care have changed (ICP > 20 vs ICP > 22).
Conclusion
Although studies supported by the NINDS adopt CDEs, they fall short when applied to annotating data in routine NCC and studies requiring rigorous analysis of monitored data. Our work will be valuable in extending concepts and use of CDEs to routine care.