Artificial Neural Networks Applied to Prediction to Assess the Likelihood of Surgical Site Infection in Different Surgeries

Background: Based on data obtained from hospitals in the city of Belo Horizonte (population ~3,000,000), we evaluated relevant factors such as death, age, duration of surgery, potential for contamination and surgical site infection, plastic surgery, and craniotomy. The possibility of predicting surgical site infection (SSI) was then analyzed using pattern recognition algorithms based on MLP (multilayer perceptron). Methods: Data were collected by the hospital infection control committees (CCIHs) in hospitals in Belo Horizonte between 2016 and 2018. The noisy records were filtered, and the occurrences were analyzed. Finally, the predictive power of SSI of 5 types MLP was evaluated experimentally: momentum, backpropagation standard, weight decay, resilient propagation, and quick propagation. The model used 3, 5, 7, and 10 neurons in the occult layer and with resamples varied the number of records for testing (65% and 75%) and for validation (35% and 25%). Comparisons were made by measuring the AUC (area under the curve (range, 0–1). Results: From 1,096 records of craniotomy, 289 were usable for analysis. Moreover, 16% died; averaged age was 56 years (range, 40–65); mean time of surgery was 186 minutes (range, 95–250 minutes); the number of hospitalizations ranged from 1 (90.6%) to 8 (0.3%). Contamination among these cases was rated as follows: 2.7% contaminated, 23.5% potentially contaminated, 72.3% clean. The SSI rate reached 4%. The prediction process in AUCs ranged from 0.7 to 0.994. In plastic surgery, from 3,693 records, 1,099 were intact, with only 1 case of SSI and no deaths. The average age for plastic surgery was 41 years (range, 16–91); the average time of surgery was 218.5 minutes (range, 19–580 minutes); the number of hospitalizations ranged from 1 (77.4%) to 6 times (0.001%). Contamination among these cases was rated as follows: 27.90% potential contamination, 1.67% contaminated, and 0.84% infected. The prediction process ranged in AUCs from 0.2 to 0.4. Conclusions: We identified a high noise index in both surgeries due to subjectivity at the time of data collection. The profiles of each surgery in the statistical analyses were different, which was reflected in the analyzed structures. The MLP for craniotomy surgery demonstrated relevant predictive power and can guide intelligent monitoring software (available in www.sacihweb.com). However, for plastic surgeries, MLPs need more SSI samples to optimize outcomes. To optimize data collection and to enable other hospitals to use the SSI prediction tool, a mobile application was developed. Disclosures: None Funding: None