BACKGROUND
Systematic reviews depend on time-consuming extraction of data from PDFs of underlying studies. To date, automation efforts have focused on extracting from the text, and no approach has yet succeeded in fully automating ingestion of quantitative evidence. However, the majority of relevant data is generally presented in tables, and tabular structure is more amenable to automated extraction than free-text.
OBJECTIVE
The purpose of this survey is to classify the structure and format of descriptive statistics reported in tables in the comparative medical literature.
METHODS
We sampled 100 published randomized controlled trials (RCTs) from the year 2019 from PubMed; these results were imported to the AutoLit platform. Studies were excluded if they were non-clinical, non-comparative, not in English, protocol-only, or not available in full text. In AutoLit, tables reporting baseline or outcome data in all studies were characterized based on reporting practices. Measurement context, meaning the structure in which the interventions of interest, patient arm breakdown, measurement timepoints, and data element descriptions were presented, was classified based on the number of contextual pieces and metadata reported. Then, the statistic formats for reported metrics (specific instances of reporting of data elements) were classified by location and broken down into reporting strategies for continuous, dichotomous, and categorical metrics.
RESULTS
We included 78 of 100 studies, one of which (1.3%) did not report data elements in tables. The remaining 77 studies reported baseline and outcome data in 174 tables, and 97% of these tables broke down reporting by patient arms. Fifteen structures were found for the reporting of measurement context, which were broadly grouped into: 1x1 Contexts, where two pieces of context are reported total (e.g. “arms in columns, data elements in rows); 2x1 Contexts, where two pieces of context are given on row headers (e.g. timepoints in columns, arms nested in data elements on rows); 1x2 Contexts, where two pieces of context are given on column headers. 1x1 contexts were present in 57% of tables, compared to 20% for 2x1 and 15% for 1x2 (8% used unique/other stratification). Statistic formats were reported in the headers or descriptions of 84% of studies.
CONCLUSIONS
In this pilot survey, we found a high density of information in tables, but with major heterogeneity in presentation of measurement context. The highest-density studies reported both baseline and outcome measures in tables, with arm-level breakout, intervention labels and arm sizes present, and reported both the statistic formats and units. The measurement context formats presented here, broadly classified into three classes that cover 92% of studies, form a basis for understanding the frequency of different reporting styles, supporting automated detection of data format for extraction of metrics.