Developing a Pressure Injury Predictive Indicator System for Data Mining in Health Care Information Systems: A Sequential Mixed-Methods Study

OBJECTIVE: Prevalence of hospital-acquired pressure injury (PI), as a critical measurement of medical care quality, has shown an upward trend. The aim of this study was to determine the predictive indicators of potential PIs and ensure that the predictive indicators can automatically be mined from electronic medical record systems. METHODS: The methods include 2 parts. One is the modified Delphi for indicator development, including clinical health care provider interviews, literature review, research group meetings, and Delphi survey. The other is feature selection, including extracting indicators from the health care information system (HIS) by structured query language and selecting indicators using the Random Forest technique. RESULTS: A predictive indicator system (with feature extraction rules for each indicator) consisting of 3 categories and 14 indicators was constructed. The experts’ consensus was reached on all indicators (mean=4.28±0.65 to 4.94±0.23; coefficient of variation=4.63% to 17.20%; agreement rate=83.30% to 100.00%). The agreement between manual extraction and the computer’s automatic extraction was good, with a Cohen κ score of 0.64 to 1.00. The accuracy of the good parsimonious prediction model was 95.26%. CONCLUSIONS: This predictive indicator system is prepared for automatic PI prediction in the HIS. Many revisions should be conducted in further studies and practices in a real-life medical environment.