Deep Learning-based Risk Prediction Model for Postoperative Healthcare-associated Infections.
10.3881/j.issn.1000-503X.14013
- Author:
Chen SUN
1
;
Li-Jian PEI
1
;
Yue-Lun ZHANG
1
;
Yu-Guang HUANG
1
Author Information
1. Department of Anesthesiology,PUMC Hospital,CAMS and PUMC,Beijing 100730,China.
- Publication Type:Randomized Controlled Trial
- Keywords:
deep learning;
healthcare-associated infection;
perioperative management;
prediction model
- MeSH:
Cross Infection;
Deep Learning;
Delivery of Health Care;
Humans;
Postoperative Period;
Retrospective Studies
- From:
Acta Academiae Medicinae Sinicae
2022;44(1):9-16
- CountryChina
- Language:Chinese
-
Abstract:
Objective To develop a risk prediction model combining pre/intraoperative risk factors and intraoperative vital signs for postoperative healthcare-associated infection(HAI)based on deep learning. Methods We carried out a retrospective study based on two randomized controlled trials(NCT02715076,ChiCTR-IPR-17011099).The patients who underwent elective radical resection of advanced digestive system tumor were included in this study.The primary outcome was HAI within 30 days after surgery.Logistic regression analysis and long short-term memory(LSTM)model based on iteratively occluding sections of the input were used for feature selection.The risk prediction model for postoperative HAI was developed based on deep learning,combining the selected pre/intraoperative risk factors and intraoperative vital signs,and was evaluated by comparison with other models.Finally,we adopted the simulated annealing algorithm to simulatively adjust the vital signs during surgery,trying to explore the adjustment system that can reduce the risk of HAI. Results A total of 839 patients were included in this study,of which 112(13.3%)developed HAI within 30 days after surgery.The selected pre/intraoperative risk factors included neoadjuvant chemotherapy,parenteral nutrition,esophagectomy,gastrectomy,colorectal resection,pancreatoduodenectomy,hepatic resection,intraoperative blood loss>500 ml,and anesthesia time>4 h.The intraoperative vital signs significantly associated with HAI were in an order of heart rate>core body temperature>systolic blood pressure>diastolic blood pressure.Compared with multivariable Logistic regression model,random forest model,and LSTM model including vital signs only,this deep learning-based prediction model performed best(ACC=0.733,F1=0.237,AUC=0.728).The simulation via simulated annealing algorithm reduced the incidence of postoperative HAI.Moreover,the incidence decreased most in the case of reducing the initial annealing temperature and choosing the last 20% of surgery procedure. Conclusions This study developed a risk prediction model for postoperative HAI based on deep learning,which combined pre/intraoperative risk factors and intraoperative basic vital signs.Using simulated annealing algorithm to adjust intraoperative vital signs could reduce the incidence of postoperative HAI to some extent.
