BACKGROUND
In the United States, State laws require birth certificates to be completed for all births; and federal law mandates national collection and publication of births and other vital statistics data. National Center for Health Statistics (NCHS) has published the key statistics of birth data over the years. These data files, from as early as the 1970s, have been released and made publicly available. There are about 3 million new births each year, and every birth is a record in the data set described by hundreds of variables. The total data cover more than half of the current US population, making it an invaluable resource to study and examine birth epidemiology. Using such big data, researchers can ask interesting questions and study longitudinal patterns, for example, the impact of mother's drinking status to infertility in metropolitans in the last decade, or the education level of the biological father to the c-sections over the years.
However, existing published data sets cannot directly support these research questions as there are adjustments to the variables and their categories, which makes these individually published data files fragmented. The information contained in the published data files is highly diverse, containing hundreds of variables each year. Besides minor adjustments like renaming and increasing variable categories, some major updates significantly changed the fields of statistics (including removal, addition, and modification of the variables), making the published data disconnected and ambiguous to use over multiple years. Researchers have previously reconstructed features to study temporal patterns, but the scale is limited (focusing only on a few variables of interest). Many have reinvented the wheels, and such reconstructions lack consistency as different researchers might use different criteria to harmonize variables, leading to inconsistent findings and limiting the reproducibility of research. There is no systematic effort to combine about five decades of data files into a database that includes every variable that has ever been released by NCHS.
OBJECTIVE
To utilize machine learning techniques to combine the United States (US) natality data for the last five decades, with changing variables and factors, into a consistent database.
METHODS
We developed a feasible and efficient deep-learning-based framework to harmonize data sets of live births in the US from 1970 to 2018. We constructed a graph based on the property and elements of databases including variables and conducted a graph convolutional network (GCN) on the graph to learn the graph embeddings for nodes where the learned embeddings implied the similarity of variables. We devised a novel loss function with a slack margin and a banlist mechanism (for a random walk) to learn the desired structure (two nodes sharing more information were more similar to each other.). We developed an active learning mechanism to conduct the harmonization.
RESULTS
We harmonized historical US birth data and resolved conflicts in ambiguous terms. From a total of 9,321 variables (i.e., 783 stemmed variables, from 1970 to 2018) we applied our model iteratively together with human review, obtaining 323 hyperchains of variables. Hyperchains for harmonization were composed of 201 stemmed variable pairs when considering any pairs of different stemmed variables changed over years. During the harmonization, the first round of our model provided 305 candidates stemmed variable pairs (based on the top-20 most similar variables of each variable based on the learned embeddings of variables) and achieved recall and precision of 87.56%, 57.70%, respectively.
CONCLUSIONS
Our harmonized graph neural network (HGNN) method provides a feasible and efficient way to connect relevant databases at a meta-level. Adapting to databases' property and characteristics, HGNN can learn patterns and search relations globally, which is powerful to discover the similarity between variables among databases. Smart utilization of machine learning can significantly reduce the manual effort in database harmonization and integration of fragmented data into useful databases for future research.
A Data Adaptive Model for Retail Sales of Electricity
