Fig 1.
(A) Development—Two BCDs were developed at IH and UPMC, respectively. At each site, a rule-based parser was manually built by a knowledge engineer based on expert-annotated sample notes. A Bayesian network classifier was machine-learned from a local training set. (B) Test—Test datasets were created using local encounters (continuous arrows) or non-local encounters (dashed arrows) to evaluate both local performance and transferability. Bayesian network classifier’s abilities to discriminate a case of (1) influenza from non-influenza and (2) influenza from non-influenza influenza-like illness (NI-ILI) were evaluated. Not shown—an algorithm limited encounter data included in the test dataset based on time since registration.
Table 1.
Summary of training and test datasets.
Table 2.
Transferability studies.
Table 3.
Seven factors affecting the performance of a Bayesian case detection system.
Table 4.
Clinical findings included in the four Bayesian network classifiers.
Fig 2.
Four Bayesian network classifiers developed using datasets distinguished by data resources and NLP parsers.
GeNIe visualization [62].
Table 5.
Performance and transferability of the influenza detection systems.
Table 6.
Information delay of all ED encounters between June 1, 2010 and May 31, 2011 at UPMC and IH.
Fig 3.
AUCs of BCDIH and BCDUPMC for discriminating between influenza and non-influenza over different time delays.
Fig 4.
AUCs of BCDIH and BCDUPMC for discriminating between influenza and NI-ILI over different time delays.