To our knowledge, the retrospective study firstly proposed a simple prognostic model (Broaln-MEWS) combined age group, BMI group, MEWS, RDW group, NLR, lac, osmolarity group to predict 28-day mortality in critically ill patients that was internal validated. The major finding of this study was that AUROC showed that prediction efficiency of 28-day prognosis of Broaln-MEWS was higher than that of the traditional MEWS, NLR, RDW, lac, or osmolarity alone and slightly lower than APACHE-II. But Broaln-MEWS was not inferior to APACHE-II score, statistically.
For ICU patients with features of complicate disease spectrum, longer in-hospital stay and higher medical resource consumption rates, early identification of the death risk of critical illness in ICU with prognostic scoring systems was important for the timely and effective management and intervention. Although there were many scoring systems of critical illness used to distil the complexity into a single measure to quantitate survival probabilities in current clinical practice, we should not ignore the drawbacks and flaws of individual systems. For instance, some assessment tools need many blood tests or scoring items to fill resulting in more time to complete causing delayed interventions or increasing the financial burden of patients. Thus, fast, convenient and cheap evaluation tools were more attractive in clinical practice.
Our study used retrospectively collected variables to predict the development of 28d-mortality in critically ill patients. The selected variables were selected from the previously literature and used in previously ICU risk assessment models. In our study, we demonstrated that non-survivors tended to be older, had lower levels of BMI, and had higher MEWS, RDW, NLR, lac, osmolarity, which indicated that age, BMI, MEWS, NLR, RDW, lac, and osmolarity values may serve as potential prognostic markers in critical illness. In these indexes, MEWS was developed as practical tools that could rapidly and effectively estimate clinical death risk by using only 5 simple and basic physiologic parameters without increasing economic burden of patients, because these parameters could be acquired by electronic medical record, automatically. According to previous observational studies, Moon et al. reported that the introduction of MEWS charts significantly reduced the number of in-hospital cardiac arrest calls (2% vs. 3%; p = 0.004) and in-hospital mortality rates(42% vs. 52%;p = 0.05) [18]. In a study to predict 28-day mortality rate of ICU patients with severe sepsis/septic shock, MEWS was associated with the 28-day mortality rate (OR, 1.462; 95%CI, 1.122 to 1.905; p = 0.005) [19], which was consistent with our study (OR, 1.250; 95%CI, 1.232 to 1.269; p = 0.000). However, some study also found that MEWS has a limited ability to estimate sudden deterioration in patients like cardiac shock [20]. Therefore, MEWS alone to predict mortality rate of critical ill patients requires further investigation. As was well known, sepsis was recognized as a global health problem in ICU. The proportion of ICU-acquired sepsis was 24.4% and the mortality in hospitalized sepsis patients remains very high at 25–30% [21]. Whether patients have sepsis was an important factor affecting the mortality of ICU patients. NLR, as immune-related biomarker, was found to serve as a convenient prognostic marker in septic patients. In a study to predict 28d-mortality in septic patients, Liu et al. reported NLR was associated with the 28-day mortality rate (OR, 1.340; 95%CI, 1.253 to 1.434; p < 0.001) [22]. However, in our study, the OR value was 1.047(95%CI, 1.041 to 1.053; p = 0.000). The reason for this difference was that Liu et al selected patients with sepsis, and the population in our study was all ICU patients, not limited to sepsis. Lac had been usually associated with mortality in different groups of critically ill patients, as seen in cardiogenic, hypovolemic, and septic shock etc. Relative hyperlactatemia (1.36 to 2.00mmol/L) within the first 24 hours of ICU admission is an independent predictor of hospital and ICU mortality in critically ill patients [16]. RDW represented an index of the variability in size of circulating erythrocytes. The normal range of RDW is 11.0%-14.5%. There were few clinical scenarios resulting in an RDW less than 11.0%, but many disease processes elevating the RDW above 14.5% [23]. A series of studies have demonstrated that RDW shows the predictive value of mortality in patients with heart failure, septic shock, acute respiratory distress syndrome, etc. [14, 24, 25]. In our study, RDW with a threshold of 14.5% showed the strong correlation with the mortality of critical ill patients (OR, 1.762; 95%CI, 1.613 to 1.925; p = 0.000). Our outcome was similar to previous study by Heidi S Bazick et al. (30d mortality RDW 14.7–15.8% OR, 1.69; 95% CI, 1.52–1.86; p < 0.001; >15.8% OR, 2.61; 95% CI, 2.37–2.86; p < 0.001) [23]. In addition, osmolarity, age, and BMI were associated with poor clinical outcomes of critical illness. Osmolarity with thresholds at 300 mmol/L was associated with increased mortality in critically ill patients with cardiac, cerebral, vascular and gastrointestinal admission diagnoses [26], which was consistent with our study (OR, 1.34; 95%CI, 1.252 to 1.464; p = 0.000). In a nationwide ICU mortality study of Poland, younger patient age was associated with ICU survival [3]. Zhigang Xue etc. report patients with BMI < 18.5 kg/m2 had significantly higher ICU mortality (OR:1.92; 95% CI:1.84–2.01) in Asian hospitalized patients [27]. Our study also revealed that age with thresholds at 60 year and BMI with thresholds at 18.5 and 28 was associated with increased mortality in critically ill patients.
Next, we investigated the factors that independently predicted the 28d mortality in critical illness. Our data showed that the age group, BMI group, MEWS, RDW group, NLR, lac and osmolarity group were independent predictors of the 28-day mortality by using logistic regression analysis. Furthermore, the ROC curves were used to evaluate the predictive power of each of the above independent factors for 28-day mortality of critical illness. It was noticed that APACHE-II had the largest AUC value (0.747), followed by MEW (0.669), RDW (0.634), osmolarity (0.592), NLR (0.591), and lac (0.580) as a single parameter. These outcomes indicated that the predictive power of MEWS, RDW, NLR, lac, or osmolarity was limited and inferior to APACHE-II. For disease complexity and heterogeneity in critical illness, combining different indexes can more accurately reflect the prognosis of ICU patients. For example, the addition of RDW to APACHE-III improved its mortality prediction marginally. Adding RDW to APACHE-III increased AUROC (from 0.9586 to 0.9613) [28]. In our study, we used age, MEWS, NLR, RDW, lac, and osmolarity to predict 28d-mortality. Notably, a significant increasement of AUC (0.741) was found after we combined these parameters as a composite index. The major reasons for this may be related to differences in these indexes reflecting six dimensions of ICU patients from: age to aging degree, BMI to nutritional status, MEWS to general condition, lac to microcirculation, NLR to sepsis and osmolarity to internal environment.
Moreover, we used Brier score to assess model accuracy. In the original work, Broaln-MEWS had a smallest Brier score of 0.103 with comparison of APACHE-II (0.106), MEW (0.113), RDW (0.116), NLR (0.118), lac (0.115) and osmolarity (0.114), indicating that Broaln-MEWS had better accuracy in prediction at an individual level. Additionally, we calculated H/L C-statistic to assess acceptable agreement between observed ICU mortality and actual ICU mortality. Of these proposed models, only Broaln-MEWS and APACHE-II had discrimination greater than 0.70, calibration was adequate (p = 0.291 and p = 0.659 > 0.05), suggesting the assignment of the correct probability at all levels of predicted risk. Finally, the Broaln-MEWS model provided stable evaluation with excellent calibration assessed by use of a validation group (AUROC: 0.744 p = 0.000; H/L C-statistic 11.37 p = 0.182).
Our study had some strengths. First, as far as we know, our work firstly demonstrated the enhanced prognostic prediction value of combination with age, MEWS, NLR, RDW, lac, and osmolarity. Second, the sample size in our study was relatively large to reduce selection bias. Furthermore, we performed different probability models to evaluate Broaln-MEWS model in order to ensure the scientific nature and credibility of the results. Third, these parameters were objective and easily accessible laboratory widely available to clinicians. For instance, RDW and NLR were routinely reported as part of the complete blood count. Forth, it was easy and not time-consuming to calculate 28d-mortality by nomogram graph provided in our study.
Nevertheless, our data were collected retrospectively by MIMIC-III database. It was important to recognize the limitations of our study. First, because our work was a single institutional retrospective study with MIMIC-III database, it was difficult to extrapolate the findings to other hospitals. An external validation in cohorts of other countries was warranted before the conclusion may be generalize. Second, for data collection from MIMIC-III database was incomplete and contained many inaccurate data elements, the bias potential could not be excluded.