The role of antibacterial pattern of resistance on mortality associated with catheter related ICU acquired infections; a machine learning-derived analysis on the results of 8 years of surveillance

ICU acquired infections are among the most common nosocomial infections. The pattern of antibacterial resistance is different according to regional and even ICU related factors. The infections caused by more resistant bacteria including pan drug-resistance (PDR), extensively drug-resistance (XDR) and multidrug-resistance (MDR) are associated with higher rates of mortality, reportedly. Machine learning (ML) algorithms can be useful for prediction of mortality in ICU settings. We aimed to examine different ML models to find the relation between the antibacterial resistance pattern of major ICU acquired infections with mortality in patients with ventilator associated pneumonia (VAP), central line associated blood stream infection (CLABSI) and catheter associated urinary tract infection (CAUTI). Data were collected prospectively from February 2014 to June 2021 in 9 adult medical and surgical ICU's in Nemazee Hospital, Shiraz, Iran. Every patient with any episode of VAP, CLABSI and CAUTI was recruited. The causative microorganisms were categorized to PDR, XDR, MDR, usual and fungal according to the laboratory culture results. We employed the random forest (RF), gradient boosting machine (GBM) and the traditional logistic regression (LR) method to find any relation between the resistance pattern of the involved microorganism with ICU mortality. In addition, the performance of the three developed models was evaluated and compared. The records from 968 patients were used for model training and test. Amongst them, 317 (32.7%) patients had in-ICU death. The average AUROCs of the RF, GBM and LR models were 0.80, 0.79 and 0.76 respectively. The sensitivity and specificity of the RF, GBM, and LR were (0.74, 0.73), (0.82, 0.61) and (0.67, 0.77). The Gini index showed that age, length of ICU-stay, ventilator free days, duration of having central line and urinary catheters and ICU bed number had more significant relation with mortality than antimicrobial resistance category of the causative agent. The calibration plots showed that the RF had better calibration than GBM and LR models. The machine learning-based models developed in this study had good performance. The pattern of antibacterial resistance was not a major determinant of ICU mortality in this patient population with ICU acquired infections. Availability of appropriate antibiotics and early and proper medical interventions may have decreased the impact of microbial resistance pattern in ICU mortality. These find