We used ML-based approaches to develop a highly accurate model in order to predict septic shock in patients with OAPN-US. To the best of our knowledge, this is the first study to apply ML for the prediction of septic shock. In particular, previous studies have reported PCT, ALC, and INR levels as biomarkers; however, these variables have not been combined to estimate their predictive performance. Notably, these three variables are generally included in routine investigations. The present study included most of the factors that were identified to be significant in the relevant previous studies [1–10, 20–25]. However, we excluded lactate and ABGA levels from the analysis as they were already included in the Sepsis-3 diagnostic criteria [19].
PCT is a precursor of a hormone that plays a role in calcium metabolism. Although it is elevated in cases of major trauma, elective surgery, and severe burns as well as in some cancers [26], it is widely used as a biomarker to detect sepsis and septic shock early and guide treatment [27]. According to a systematic review and meta-analysis that included approximately 1300 patients from 9 studies, PCT demonstrated a higher diagnostic accuracy than CRP for sepsis [28]. According to Brodska et al., higher levels of PCT may be more useful in urinary tract infections as patients with sepsis and positive blood cultures responded more strongly to gram-negative than to gram-positive bacteria [29]. Moreover, it was reported that PCT levels can help distinguish between gram-negative and gram-positive sepsis, indicating that measurement of PCT levels differs from that of CRP levels [29]. As shown in Table 2, most urinary tract infections were caused by gram-negative bacteria in the present study. However, the utility of PCT levels as a biomarker is limited in predicting mortality. Andriolo et al. [27] performed a meta-analysis for adults with sepsis, severe sepsis, or septic shock using the Sepsis-2 criteria. Based on the comparison of PCT with standard care or CRP for predicting mortality at 28 days and ICU discharge, they concluded that PCT offers low-quality evidence. In contrast, PCT was found to be helpful in predicting mortality at hospital discharge. Furthermore, the duration of antimicrobial therapy was not related to PCT levels.
ALC of < 1,000 cells/mm3 is generally considered to indicate lymphocytopenia [30]. In the present study, 33 of 37 (89.2%) and 22 of 44 (50.0%) patients in the SS and NSS groups met the abovementioned criterion, respectively. Moreover, similar to PCT levels, lymphocytopenia can occur during sepsis, burns, trauma, general anesthesia, and major surgeries according to a previous study [31]. It has been suggested that lymphocytopenia plays an essential role in severe infections, such as sepsis and septic shock. According to Tulzo et al. [32] and Valet et al. [33], patients with septic shock exhibit higher levels of lymphocyte apoptosis, with decreasing counts of CD4+, CD8+ T, and NK cell subpopulations, which further favor immunosuppressive conditions. A study by Sheikh Motahar Vahedi et al. [34] reported that lymphocytopenia was independently associated with higher 28-day mortality rates, and patients with lymphocytopenia were older than individuals in the control group; moreover, such patients required more ICU admissions as well as had a significantly higher rate of 28-day septic shock and readmission due to sepsis and a higher SOFA score. In an analysis of 58,260 cases of hospital admissions, Andreu-Ballester et al. [31] demonstrated that lymphocytopenia is a predictor of the severity of illness and mortality risk. Consistent with the findings of these studies, our study revealed that lymphocytopenia should be included in OAPN-US’s prediction model for septic shock.
The World Health Organization developed INR to standardize the variations observed in PT. With the addition of thromboplastin extracts and ionized calcium to the plasma, PT can measure the extrinsic coagulation pathway [35]. Notably, disseminated intravascular coagulopathy is associated with prolonged PT and INR combined with thrombocytopenia [36]. In cases of septic shock, thrombocytopenia is associated with poor outcomes, including short overall survival after ICU admission, prolonged length of stay, long duration of organ support, additional bleeding events, and even mortality [37–40]. Patients with thrombocytopenia experience poor outcomes owing to endothelial dysfunction, coagulopathy, hemodilution, and altered thrombopoiesis [41, 42]. Although the use of INR alone as a prognostic factor has not been widely suggested, two recent studies have suggested that using INR alone is effective. According to Ling et al. [43], INR was a significant predictor of mortality in patients with sepsis and necrotizing fasciitis. Furthermore, Zhang et al. [44] suggested using INR to evaluate nonpulmonary infectious sepsis in adults at the early stages of the disease process. Similarly, our study concluded that the standardized ratio INR is a better indicator of the severity of septic shock than platelet count or PT.
The results of our study are consistent with those of previous studies. Notably, two studies used the Sepsis-3 criteria to predict sepsis and septic shock in patients with OAPN-US. In 2020, Tambo et al. [9] performed multiple logistic regression analyses of 50 patients with OAPN-US, including 11 patients with sepsis. Although the primary endpoint of their study was sepsis, PCT and presepsin were identified as the most significant predictors of sepsis. In the study by Baboudjian et al. [10], multiple logistic regression analyses were performed for 110 patients with OAPN-US, including 39 patients with septic shock. The results of their study also indicated that PCT, CRP, and DM were the major predictors of septic shock. In the present study, PCT levels demonstrated excellent predictive performance, with an AUC of 0.91. Furthermore, based on the results of multivariate logistic regression analyses, both the previous studies concluded that PCT level is the most important predictor of septic shock.
The results of our study were also consistent with studies conducted in accordance with the Sepsis-2 criteria. For cases of OAPN with or without US, Kamei et al. [2] suggested the use of thrombocytopenia and a positive blood culture as predictors of septic shock. Moreover, in a study involving 69 patients with OAPN-US, including 23 patients with septic shock, Tambo et al. [3] suggested that thrombocytopenia and low serum albumin levels were additional predictors of septic shock. Furthermore, according to Ko et al. [4], thrombocytopenia and increased PCT levels were identified as predictors of septic shock based on the examination of 49 patients, including 15 patients with septic shock. In addition, based on the data of 69 patients with OAPN-US, including 25 patients with septic shock, Kakinoki et al. [5] reported serum albumin levels as a predictor of septic shock. Moreover, leukocytosis has been suggested as a predictor of septic shock by Srougi et al. [22] in a study involving 52 patients with OAPN-US, of whom, 6 suffered from septic shock. Notably, DM and CRP were determined as the most important predictors of septic shock by Yamamichi et al. [6]. A prospective study by Kozyrakis et al. [23] involved 62 patients with OAPN-US, of whom, 20 had sepsis and 6 had septic shock. Their results indicated a higher age, the existence of multiresistant strains, and a higher creatinine level as risk factors for sepsis. They also suggested that DM is a risk factor for longer hospitalizations. Lee et al. [8] identified urinary tract calculi, leukocytosis, elevated CRP levels, and low albumin levels as predictors of septic shock. In addition, a decreased ESR and the absence of hypertension were considered as potential predictors. Despite the differences in detailed diagnostic criteria between the abovementioned studies and our study, the conclusions were similar. In addition, most of the factors that were suggested as risk factors or predictors in previous studies were considered in our study.
This study has some limitations. First, this was a retrospective study conducted in a single institution. Second, we could include only a few patients. Moreover, for patients with mild infections, such as uncomplicated acute pyelonephritis, analyses of PCT levels are not performed at our center. As this was not a prospective study, PCT levels could not be analyzed for all patients with OAPN-US. Thus, patients who did not undergo analysis for PCT levels were excluded from the study as multiple previous studies have reported the significance of PCT. Accordingly, > 50% of patients were excluded from the study, which may have led to bias. Finally, we did not perform imputations because it was not appropriate to impute important values.