Prediction of Surgical Risk in General Surgeries: Process Optimization Through Support Vector Machine (SVM) Algorithm

Abstract

Background: In 5 hospitals in Belo Horizonte (population, 3 million) between July 2016 and June 2018, a survey was performed regarding surgical site infection (SSI). We statistically evaluated SSI incidents and optimized the power to predict SSI through pattern recognition algorithms based on support vector machines (SVMs). Methods: Data were collected on SSIs at 5 different hospitals. The hospital infection control committees (CCIHs) of the hospitals collected all data used in the analysis during their routine SSI surveillance procedures; these data were sent to the NOIS (Nosocomial Infection Study) Project. NOIS uses SACIH software (an automated hospital infection control system) to collect data from hospitals that participate voluntarily in the project. In the NOIS, 3 procedures were performed: (1) a treatment of the database collected for use of intact samples; (2) a statistical analysis on the profile of the hospitals collected; and (3) an assessment of the predictive power of SVM with a nonlinear separation process varying in configurations including kernel function (Laplace, Radial Basis, Hyperbolic Tangent and Bessel) and the k-fold cross-validation–based resampling process (ie, the use of data varied according to the amount of folders that cross and combine the evaluated data, being k = 3, 5, 6, 7, and 10). The data were compared by measuring the area under the curve (AUC; range, 0–1) for each of the configurations. Results: From 13,383 records, 7,565 were usable, and SSI incidence was 2.0%. Most patients were aged 35–62 years; the average duration of surgery was 101 minutes, but 76% of surgeries lasted >2 hours. The mean hospital length of stay without SSI was 4 days versus 17 days for the SSI cases. The survey data showed that even with a low number of SSI cases, the prediction rate for this specific surgery was 0.74, which was 14% higher than the rate reported in the literature. Conclusions: Despite the high noise index of the database, it was possible to sample relevant data for the evaluation of general surgery patients. For the predictive process, our results were >0.50 and were 14% better than those reported in the literature. However, the database requires more SSI case samples because only 2% of positive samples unbalanced the database. To optimize data collection and to enable other hospitals to use the SSI prediction tool, a mobile application was developed (available at www.sacihweb.com).

Funding: None

Disclosures: None