SSI cases. Although the sources of data were all electronic, the
process of screening and review was carried out manually by
trained personnel. The algorithm performed well but the processing
of data required an increased amount of manual input
compared with other ESSs reported in the literature. This system,
did, however, reduce the number of charts required for
manual review by 90.5% (NPV ¼ 100%).
Shaklee et al. evaluated the performance of administrative
data for the detection of CDI.49 ICD-9-CM coding data and
billing data for testing and treatment were compared with
positive microbiological results. ICD-9-CM codes alone
demonstrated the best performance in terms of sensitivity,
illustrating that ICD-9-CM codes for CDI were reliable and accurate
in identifying hospitalized children with CDI.
Wright et al. used information contained within the electronic
medical record to identify patients with invasive devices.
51 Although this system was not developed to monitor
HCAI, the information generated is imperative to the calculation
of rates of HCAI and patients at risk of developing devicerelated
infections. To assess the performance of the automated
system, medical records from a random sample of patients
were reviewed to determine the presence of indwelling
devices. The automated system was successful in the identification
of all three devices surveyed (Table IV), and the data
collection process required much less time than manual data
collection, resulting in an estimated reduction of 142 h/year.
Single-source electronic surveillance systems
All three methods using a single data source used Natural
Language Processing (NLP) techniques for HCAI identification.
52e54 Haas et al. compared the performance of traditional
surveillance for nosocomial pneumonia and
computerized surveillance of chest x-rays using an automated
detection system based on NLP.52 NLP converts electronic
narrative into coded descriptions appropriate for the automated
rule-based system. This method had a low PPV (7.9%)
but a very high NPV (>99%), so could be used to target the
investigation of patients with suspicious chest x-rays rather
than having to assess all patients. Penz et al. found that
spelling errors and abbreviations were a clear source of error in
NLP-based methods, and appreciated that this needed to be
addressed.53
Proux et al. recognized that tools to automate reporting are
necessary as information is spread across various databases
containing ever-evolving information, which makes data difficult
to obtain.54 The aim of their study was therefore to produce
a tool capable of detecting HCAI automatically from
patient records using NLP. A series of text-based rules were
applied to the patient discharge summary, which meant that
their method relied solely on events documented in medical
records.