Establishment and Veri � cation On A Death Risk Model of Sepsis Patients Within 30 Days


 Background/Objective:

To establish and validate an individualized nomogram to predict the probability of death within 30 days in patients with sepsis would help clinical physicians to make correct decision.
Methods

We collected data of 1,205 patients with sepsis. These included 16 indexes like age and blood, randomly assigned to the modeling and verification groups. In the modeling group, the independent risk factors related to death within 30 days were analyzed. Besides, a nomogram was established to draw the receiver-operating characteristic (ROC) curve of the subjects. Subsequently, the discriminant ability of the model was evaluated by the area under the ROC curve (AUC). Then, a calibration chart and Hosmer-Lemeshow test were employed to evaluate the calibration degree of the model, and the Decline Curve Analysis (DCA) test was used to evaluate the clinical effect of the model.
Results

The different independent risk factors related to the death of sepsis patients within 30 days included pro-brain natriuretic peptide (pro.bnp), albumin, lactic acid (lac), oxygenation index, mean arterial pressure (map), and hematocrit (hct). The AUC of the modeling and verification groups were 0.815 and 0.806, respectively. Moreover, the P-values of the Hosmer-Lemeshow test in the two groups were 0.391 and 0.100, respectively, and the DCA curves of the two groups were both above the two extreme curves.
Conclusion

Our model presents good significance for predicting the death of sepsis patients within 30 days. Therefore, there is a need to implement this model in clinical practice, as prompt prediction could help tailor treatment regimens and enhance survival outcomes.


Introduction
Sepsis has a high fatality rate, leading to increased costs in healthcare [1][2][3][4],and identifying severe sepsis patients with a predicted high risk of death and early interventions are the key strategies for improving prognosis [5]. Clinically, to predict outcome of patients with sepsis, various severity scores have been widely used. However, these methods have shortcomings. For example, the APACHEII score is widely applied to assess the prognosis of critically ill sepsis patients but lacks pertinence [6], studies have pointed out that APACHE II score underestimates the risk of death in patients with sepsis in ICU [7]. studies [8,9]. On the other hand, the SOFA score describes the development of multiple organ dysfunction syndromes, but also shows defects when evaluating prognosis [10]. To date, there is a lack of evaluation methods for the short-term prognosis of sepsis patients. Therefore, this study aims to establish a risk prediction model for sepsis patients to predict their death risk within 30 days and only intervene in the high-risk patients to reduce the mortality rate.

Materials And Methods
Subjects: This study included 1,205 patients diagnosed with sepsis from June 1, 2013, to September 1, 2020. Our inclusion criteria were as follows: 1. A diagnosed infection; 2. Infection caused the SOFA score to increase by 2 points or more. The exclusion criteria were: 1. Less than 18 years old; 2. Patients with leukemia, lymphoma, and end-stage tumor; 3. Patients with uncertain prognosis indicators; 4. Patients with some missing sepsis indicators.

Observation indexes
Basic patient information (gender and age) and blood sample indexes, including oxygenation index, high sensitivity -C reactive protein (crp), creatinine (cr), emergency procalcitonin (pct), activated partial thromboplastin time (aptt), albumin, prothrombin time (pt), pro.bnp, pH value, lactic acid (lac), hematocrit (hct), platelet (plt), and the mean arterial pressure and heart rate at the time of hospital admission, were recorded for the rst time within 24 hours after hospital admission. The international common unit at present was taken as the index unit.

Ethics statement
This study was approved by the Ethics Committee of Dongyang People's Hospital, and the written informed consents of all patients were obtained. All data were analyzed anonymously, and no personal information was recorded during sessions. This study was conducted following the principles outlined in the Helsinki Declaration and its amendments.
Methods: The patient data (in excel, pct>100 is marked as 100; pro.bnp>35000 is marked as 35000) were collected. The createDataPartition function in R statistical package was employed to randomly divide patients into modeling and veri cation groups according to a death ratio of 7:3.
In the modeling group, the visreg function evaluated whether there is a linear relationship between each continuous variable and death. Since the continuous variables lacked a linear relationship with death, they were converted into classi ed variables. Moreover, the one-factor analysis related to death was performed using the twogrps function (found in the CBCgrps package), which can automatically distinguish whether the data are per normal distribution. Thus, the correlation test method was selected. Multiple collinearity tests were conducted on meaningful variables in the single-factor analysis, and variance expansion coe cient and VIFs values were adopted to enhance interpretation. VIFs 10 indicated no multiple collinearities among variables. Therefore, multivariate analysis was included after excluding multiple collinearities.
Signi cant variables in multivariate analysis were used to establish prediction models, which were presented as a nomogram. The nomogram could directly show the relationship between each variable and death and calculate the risk of death individually [11].
The model was evaluated from three perspectives: discrimination, correction, and clinical signi cance.
AUC evaluated the discrimination of the model. The AUC value between 0.5 and 1.0 indicated that the model was meaningful. The closer the AUC value is to 1, the better the discrimination ability of the model. Speci cally, the AUC>0.75 demonstrated that the model had a good discrimination ability [12].
The calibration of the model was evaluated by calibration chart and Hosmer-Lemeshow chi-square test. The bootstrap method was applied in the modeling and veri cation groups, proving that the model had high stability. The tting curve showed a high overlap with the standard curve, indicating a good calibration degree. The Hosmer-Lemeshow test, rst discovered by Hoslem in R language, was employed, and its p-value was greater than 0.05, suggesting that the model tting was good [13].
The DCA curve assessed the clinical validity of the model. Interpretation of the curve was as follows: the ordinate denoted the bene t of the evaluation model, and the abscissa denoted the risk value of illness; the model curve was the curve represented by the model, with the 'All' and 'None' curves taken as reference. It was considered that the further the model curve is from the two extreme curves, the greater the clinical signi cance [14,15].

Statistical analysis
In this study, the continuity of the measured data, which followed a normal distribution, was analyzed by T-test and expressed as mean±standard deviation. Data that were not normally distributed were analyzed by the Wilcoxon rank-sum test (sample size was less than 5000) and expressed as quartile. Finally, classi cation variables were analyzed by chi-square test and expressed as a percentage, and the above statistical analysis was completed in R software.

Results
Herein, 1,205 patients were included. The modeling group included 844 cases with 124 death, whereas 361 cases were included in the veri cation group with 67 death. Thus, the mortality was about 15.8%. In the modeling group, univariate analysis showed that 11 indicators like PCT, MAP, and creatinine were related to prognosis (Table 1).  Here, brain pro-natriuretic peptide, lactic acid, albumin, oxygenation index, mean arterial pressure, and hematocrit were all independent risk factors ( Table 2). Subsequently, the independent risk factors were selected to establish a nomogram (Figure 1). The speci c values of each independent risk factor matched the corresponding scores. Then, the scores were summed to obtain the total score, indicating the corresponding death risk. For example, the patient's albumin content was 27.3g/L; oxygenation index was 300; the mean arterial pressure was 70-105mmhg; pro. bnp was 10,001-20,000pg/ml; the lactic acid value was 7.3mmol/L; hct was 0.299. Each item had a corresponding score in β (XM) terms of the nomogram, marked by a red dot. By calculation, the Total score was 2.36, and in Pr (death), the corresponding mortality rate was 21.8% (marked by a red arrow).
Interpreting the model's discrimination, the goodness of t, and clinical validity were performed using the ROC curve to perform the discrimination of models (Figure 2A). Its AUC value was 0.815, showing a good discrimination ability for the model. The optimum threshold of this model was 0.117, with a speci city of 69.6% and a sensitivity of 77.4%. The calibration chart ( Figure 2B) and Hosmer-Lemeshow test were used to assess the goodness of tting, and the p-value of the Hosmer-Lemeshow test was 0.391, suggesting good tting. Finally, the DCA curve ( Figure 2C) was used to assess clinical effectiveness. The DCA curves of both the modeling and veri cation groups were above the two extreme curves, indicating that good clinical signi cance.
Lastly, the veri cation population was interpreted: In the modeling group, discrimination ( Figure 3A), the goodness of tting ( Figure 3B), and clinical effectiveness ( Figure 3C) were interpreted, and the AUC value of the veri cation group was 0.806. The P-value of the calibrated Hosmer-Lemeshow test was 0.100, and the overlap between the tting curve and the standard curve was high, indicating a good prediction effect and good tting. Hence, the DCA curve suggested a good clinical e cacy.

Discussion
This study has established a nomogram. It exhibits different indexes as the independent risk factors causing the death of sepsis patients within 30 days. These include the pro. bnp, lactic acid, albumin, hematocrit, oxygenation index for the rst time after hospital admission, and the average arterial pressure at admission. This model will help clinicians to analyze the prognosis of sepsis patients and formulate possible intervention measures. Notably, the model was evaluated based on different perspectives, including discrimination, calibration, and clinical effectiveness, and all their results suggest that the model has good signi cance Sepsis is caused by an inability of the immune system to eliminate invading pathogens and immune disorders [16,17]. It is often manifested as multiple organ dysfunction, including coagulation disorder, cardiac dysfunction, renal insu ciency, nutritional disorder, etc. Many indexes such as PCT[18],CRP[18], renal insu ciency [19], lactic acid [20][21][22], pro. bnp [23,24], albumin [25,26], and coagulation [27,28] have been reported to have a good effect on the diagnosis and prognosis evaluation of sepsis. This study presents that PCT, CRP, platelet, coagulation, and renal function indicators have less signi cance in assessing the short-term prognosis of sepsis patients. The PCT and CRP indicators are among the commonly detected indexes in patients with sepsis and are helpful to diagnose and evaluate the severity of sepsis. However, their values for prognosis evaluation are controversial. Some studies have found no signi cant correlation between the prognosis of severe patients and the PCT and CRP levels [28,29]. Abnormal coagulation is common in patients with severe sepsis. Compared to previous studies, this study indicates that the initial coagulation index was not an independent risk factor for the death of patients with sepsis. Besides, it could be related to the presently increased clinical interventions (platelet and plasma transfusion), which could timely correct coagulation functioning. Previous studies have suggested that creatinine has signi cance in evaluating the prognosis of sepsis patients [19,30]. However, our study shows that the initial creatinine value has little signi cance in evaluating prognosis. Thus, the mature and early application of the CRRT technology could be the reason for the declining renal function indicator values in assessing the prognosis of sepsis patients [31].
Lactic acid is one of the indexes of oxygen metabolism, which could gauge prognosis in sepsis and many other critical diseases [20][21][22]. Pro-bnp is an index re ecting cardiac functioning. Sepsis could trigger septic cardiomyopathy. A study has reported that the death rate of septic patients with septic cardiomyopathy complications increases [32]. Moreover, inappropriate rehydration during treatment could also increase the pro-bnp index. Whether it is caused by sepsis or improper medical treatment, the abnormal pro-bnp increase is considered one indicator for the poor prognosis of sepsis patients [33].
Albumin is an important nutritional index, and since its consumption in sepsis patients increases, malnutrition triggers poor prognosis [25,26].
The low MAP suggests that sepsis patients are in a state of low perfusion shock, indicating that the disease has entered a severe stage. Elsewhere, studies have reported that the mortality rate of septic shock could reach 33.5%-61% [34,35],, which signi cantly increases the mortality rate of patients with sepsis.
Many studies have explored the relationship between sepsis and acute respiratory distress syndrome (ARDS), and it is generally believed that sepsis combined with ARDS would increase the mortality rate [36]. Nevertheless, few studies have assessed the relationship between sepsis and oxygenation index. This study suggests that the early decline of the oxygenation index is an independent risk factor for the prognosis of sepsis. Besides, ARDS, cardiac dysfunction, and excessive uid resuscitation are the reasons for the decreasing oxygenation index and should be actively prevented and treated in clinical medicine.
Furthermore, this study suggests that hematocrit has certain signi cance in evaluating the short-term prognosis of sepsis patients. Whether the hematocrit is lower or higher than the normal level, it eventually affects the prognosis. Hence, correcting anemia and changing blood concentration could improve the prognosis of sepsis patients.
Unlike previous studies, the independent risk factors related to the death of patients with sepsis within 30 days have been found in this study and presented using a nomogram.
Clinically, various scoring systems had been widely used in the patients with sepsis, but the ability of those scoring systems is insu cient in accurately and reliably predicting mortality in the sepsis patient population. Arabi et al. evaluated four scoring systems in ICU patients with sepsis, reporting poor calibration for all four scores [9].Compared to the old scoring system,the predictive factors included in this study are objective and simple,and the model has good signi cance in discrimination, calibration and clinical validity.
Thus, this study could provide a short-term prognosis of sepsis patients in a more intuitive, concrete, and vivid manner when compared with previous studies. Besides, it comprehensively evaluates the model, suggesting that the model is helpful for clinical decision-making.
The limitation of this study (1) in a retrospective study, selection bias cannot be avoided; (2) it is a single-center study, and real external veri cation data are lacking.
Declarations Figure 1 Nomogram for predicting the risk of short-term (30 day) death in patients with sepsis