Using Machine Learning to Detect Hospital-Specific Risk Factors of Surgical Site Infections

Abstract

Background: Identification of healthcare-associated infections (HAIs) is just a first step in the surveillance of HAIs. The other part is the analysis and interpretation of collected data, which should help to set up effective preventive measures targeted where they are needed the most. General risk factors of HAIs are mostly well known, but how do the environment and processes of each hospital affect risks of HAI? Can advanced methods of data analytics reveal hidden hospital-specific risk factors of surgical site infections (SSIs)? Methods: We analyzed data from electronic health records stored in the clinical information system of Hospital Jihlava, Czech Republic, with 650 beds and 7,500 surgeries performed annually. For each inpatient stay with a surgical procedure, we automatically observed almost 1,500 features that could lead to a higher incidence of SSIs. These features consist of patient demographic data, information from structured data (eg, patient diagnoses, departments, specific rooms, operating theaters, surgeons and other hospital staff participating in the surgery), and information extracted from clinical notes using natural language processing (eg, procedures, invasive devices, and comorbidities). We used a model based on survival analysis to reveal the risk factors that can increase the probability of SSI during the inpatient stay or outpatient care after discharge. Results: We automatically evaluated risk factors weekly for 4 months (July 2019–October 2019). We detected 16 distinct significant risk factors during this period—between 2 and 6 active risk factors each week. For example, patients visiting a specific department were up to 5 times more likely to develop an HAI than the rest of the patients (P < .001). Some of the risk factors revealed were significant only within a short time, and some of them occurred perpetually. When a feature became significant, it was considered an early warning of a problem that should be addressed by the infection prevention and control team. Trends in risk factors coefficients can also help in assessing the performance of the launched preventive measures. Conclusions: Advanced data analytics can effectively uncover hospital-specific risk factors affecting surgical site infections. Such systems can automatically deliver results that can be further explored and used as a basis for targeted preventive measures.

Funding: Datlowe provided support for this study.

Disclosures: Jakub Kozák reports salary from and ownership of Datlowe.