Machine learning for infection risk prediction in postoperative patients with non-mechanical ventilation and intravenous neurotargeted drugs

Abstract

Drug efficacy can be improved by understanding the effects of anesthesia on the neurovascular system. In this study, we used machine learning algorithms to predict the risk of infection in postoperative intensive care unit (ICU) patients who are on non-mechanical ventilation and are receiving hydromorphone analgesia. In this retrospective study, 130 patients were divided into high and low dose groups of hydromorphone analgesic pump patients admitted after surgery. The white blood cells (WBC) count and incidence rate of infection was significantly higher in the high hydromorphone dosage group compared to the low hydromorphone dosage groups (p < 0.05). Furthermore, significant differences in age (P = 0.006), body mass index (BMI) (P = 0.001), WBC count (P = 0.019), C-reactive protein (CRP) (P = 0.038), hydromorphone dosage (P = 0.014), and biological sex (P = 0.024) were seen between the infected and non-infected groups. The infected group also had a longer hospital stay and an extended stay in the intensive care unit compared to the non-infected group. We identified important risk factors for the development of postoperative infections by using machine learning algorithms, including hydromorphone dosage, age, biological sex, BMI, and WBC count. Logistic regression analysis was applied to incorporate these variables to construct infection prediction models and nomograms. The area under curves (AUC) of the model were 0.835, 0.747, and 0.818 in the training group, validation group, and overall pairwise column group, respectively. Therefore, we determined that hydromorphone dosage, age, biological sex, BMI, WBC count, and CRP are significant risk factors in developing postoperative infections.

Keywords: anesthesia; hydromorphone; infection; machine learning; neurovascular; post-surgical ICU.

Figure 1

Feature ranking and filtering process for Random Forest and SVM-RFE models. (A) Bar chart showing a random forest model's importance ranking of each variable. PLT, age, MAP, CRP, and HR are the top five variables identified. (B) ROC curves showing the classification ability of the random forest model. (C) The feature screening process of SVM-RFE results in the model with the lowest RMSE when 16 variables are selected. (D) ROC curves showing training, test, and overall classification performances of the SVM model.

Figure 2

SVM-RFE as well as logistic regression models are used for screening important clinical features. (A) Venn diagram showing six features associated with infection prediction. (B) PCA showing that based on these six characteristics, a better distinction can be made between infected and uninfected patients. (C–H) ROC curves showing the predictive performance of (C) age, (D) biological sex, (E) BMI, (F) hydromorphone concentration grouping, (G) WBC, and (H) CRP on infection.

Figure 3

A nomogram based on six factors is constructed and its accuracy is assessed. (A) The ROC curves for the logistic regression model constructed based on the six identified clinical factors for infection in training, validation, and overall pairwise column sets demonstrate better classification performance in all three datasets. (B) The nomogram was constructed using a logistic regression model. (C) Calibration plot showing the predicted values of the model are roughly consistent with the true labels, indicating that the model is reasonably accurate. (D) Clinical decision curve showing the prediction results in an overall pairwise column.

Figure 4

Compared between infected and non-infected patients, duration of ICU stay and hospitalization. In the infected group, the hospital stay and ICU stay were longer than in the non-infected group.