For each of the 107,205 citations in the full text dataset, the citation was compared to all of the other citations in the full text dataset.
We use two public datasets, namely, ICDAR 2013 dataset and the Street View Text dataset to evaluate the performance of our method.
Algorithm 2 (developed on the basis of the full-text dataset, which included panel-confirmed infectious disease deaths and 50% of the noninfectious disease-related deaths) was applied to the full-text dataset only.
The algorithms were applied to the remaining set of records, which included all 1995 OMI autopsied cases except for the development group set of cases (ntrc = 1,380 for the truncated dataset; nft = 732 for the full-text dataset).
We applied algorithm 1 to both the truncated and full-text datasets and applied algorithm 2 to the full-text dataset to determine whether an advantage existed in developing an algorithm that used the data from full-text instead of data that could be obtained from death certificates.
Algorithm application on the truncated dataset achieved a sensitivity similar to that achieved with the full-text dataset, and with a higher PVP, for deaths in which an infection is a cause of death rather than incidental to the death.
In addition, for the full-text dataset, sensitivity and PVP were not substantially compromised by including in the search, infections incidental to the cause of death.
When algorithm 1 was applied to the full-text dataset, it classified 937 (68%) of 1,380 deaths as infectious disease related.
Because we found a low predictive value positive (PVP) and low specificity when algorithm 1 was applied to the full-text dataset, we chose to develop algorithm 2 in a different manner.
R code is integrated in the text, and datasets are used throughout.
