| Summary: | 博士 === 國立臺灣大學 === 生醫電子與資訊學研究所 === 101 === Healthcare-associated infections (HAIs) are a major patient safety issue, and related pathogen, such as multidrug-resistant organism (MDRO) and Candida species, are causing a global crisis. These adverse events add to the burden of resource use, promote resistance to antibiotics, and contribute to patient deaths and disability.
A Web-based HAI surveillance system was developed for automatic integration, analysis, and interpretation of HAIs and related pathogens. Rule-based classification, population-based and patient-based pathogen surveillance were incorporated in the system, and control chart analysis, clustering analysis and data mining algorithm were implemented to facilitate infection control surveillance.
Electronic medical records from a 2200-bed teaching hospital in Taiwan were classified according to predefined criteria of HAIs and MDROs in rule-based classification system. The detailed information in each HAI was presented systematically to support infection control personnel decision. Comparing to infection control personnel’s review, this system has sensitivity of 98.16%, specificity of 99.93%, positive predictive value of 95.81% and negative predictive value of 99.97%. The consistency of HAIs’ time trends (R2=0.89) and department distribution (R2=1.00) between in absence and in presence of the system were also proved. The healthcare-associated bloodstream infection detection delay is significantly decreased after using this system (P<.001).
Then, the numbers of organisms in each MDRO pattern were presented graphically to describe spatial and time information in population-based pathogen surveillance system. Hierarchical clustering with 7 upper control limits (UCL) was used to detect suspicious outbreaks. The system’s performance was evaluated in three parts: HAIs and MDROs classification, outbreak detection based on vancomycin-resistant enterococcal outbreaks, and infection prediction based on candidiasis. The optimal UCL for MDRO outbreak detection was the upper 90% confidence interval (CI) using germ criterion with clustering (area under ROC curve (AUC) 0.93, 95% CI 0.91 to 0.95). The performance indicators of each UCL were statically significant higher with clustering than those without clustering in germ criterion (P < .001), patient criterion (P = .04), and incident patient criterion (P < .001).
Finally, there were 3 data mining algorithms including support vector machine, decision tree and inductive logic programming (ILP) being used for patient-based Candida infection prediction, and a generalized linear model was set as the baseline. In addition, the effect of adding background knowledge into ILP was also evaluated. The optimal Candida infection prediction model was ILP with background knowledge from specialist and computer algorithms, having F1 score of 0.437 and accuracy of 0.713. This research provided a preliminary result of applying data mining algorithms to Candida infection prediction, approving that adopting background knowledge could improve the performance of Candida infection prediction (P=.015).
This system automatically identifies HAIs and MDROs, accurately detect suspicious outbreak of MDROs based on the antimicrobial susceptibility of all clinical isolates, and effectively classifies Candida infection.
|
|---|
