Clinical Assistant Professor Lucille Packard Children's Hospital Stanford Santa Clara, California, United States
Abstract: Introduction/
Objectives: Easy availability of adult clinical and physiological waveform (PW) data, primarily in research repositories, has had a profound impact on the research community and has led to numerous clinical and methodological advances. However, pediatric patients differ from adults to such a degree that models developed for, and trained on, adult data do not generalize well to pediatric populations. No large, open repositories of high-frequency pediatric PW data are widely available for research.
Methods: PWs, vital signs, demographic and clinical data were downloaded from bedside patient monitoring systems and hospital electronic health records at an academic pediatric hospital from 06/2008-01/2017. All data was deidentified, cleaned and organized, and PW data was extracted and codified into a scalable open-source format to facilitate research (Figure 1). The dataset (WAVES) was uploaded and stored in Redivis, an open-access web-based research platform managed by Stanford University. WAVES can be queried by hospital encounter and includes PW, dates, and vital signs that can be filtered and joined before downloading to local files for analysis. Algorithms to process data for deep learning applications were designed, validated, and uploaded to the research platform.
Results: WAVES is a single-institution dataset comprising 9 years of high-frequency PW data from pediatric patients at a large academic medical center. WAVES consists of 10.6 million hours of 1 to 20 concurrent types of high-frequency PWs (Table 1). Approximately 1.5 million PW samples were collected over 50,364 unique hospital encounters across various specialized and general units (Table 2). Initial work demonstrated the suitability of the data for training deep learning models by accurately detecting hypotension with data only from electrocardiogram, plethysmography, and respiration waveforms.
Conclusions: WAVES is currently the largest pediatric-focused PW dataset available for open-access research and the second largest repository of correlated multi-channel PW data (second to the MIMIC-III clinical database). The WAVES database can enable improvements in pediatric clinical care through machine learning research on PWs from a variety of hospitalized pediatric patients and could facilitate the development of methodological and clinical innovation in the field of pediatric care.
