The complete blood count is frequently used to evaluate sepsis, with focus on the white cell count (WCC) and the presence of left shift or bandaemia1. However, an abnormal WCC is not a sensitive marker even in patients with bacteraemia2. Although bandaemia is more sensitive for identifying occult bacteraemia2, the technical need for manual cell count translates to a substantial delay in diagnosis3. The bandaemia response itself is also subject to delay and only emerges one day after clinical infection1. These confounding factors have led to a search for more effective markers to aid in the evaluation of infections.

The neutrophil-to-lymphocyte ratio (NLR) is a readily available marker derived from the CBC as a ratio of absolute or relative neutrophil and lymphocyte counts. Endogenous catecholamines and cortisol are released in response to physiological stress, causing an increase in neutrophils and a decrease in lymphocytes4,5. Additionally, lymphocyte apoptosis occurs in sepsis, leading to lymphopaenia6 and resulting in an elevated NLR. This response promptly occurs within 4 to 8 h of an acute insult7, making the NLR superior to leucocytosis or bandaemia for timely reflection of acute illness.

Studies have shown an association between NLR and patient outcomes in septic and bacteraemic patients in the Emergency Department and in the general ward8,9,10,11,12, as well as in acute coronary syndrome, acute pancreatitis and rheumatic diseases13,14,15,16,17,18,19,20,21,22. Its prognostic significance in the intensive care unit (ICU), however, remains uncertain. Similar to the NLR, the monocyte-to-lymphocyte ratio (MLR), platelet-to-lymphocyte ratio (PLR), and red cell distribution width (RDW) are described and evaluated as inflammatory biomarkers in a variety of medical conditions.

The use of NLR may also have diagnostic significance. In bacterial infections, neutrophilia and bandaemia develop, resulting in an increase in NLR. A higher NLR may indicate that an infection is bacterial rather than viral in origin.

Our study aimed to evaluate (1) the prognostic accuracy of these simple biomarkers in predicting hospital mortality compared to the Acute Physiology and Chronic Health Evaluation (APACHE) IV score and (2) their diagnostic power in differentiating pneumonia aetiologies.


Study design and data collection

This retrospective analysis was conducted from January 1, 2017, to June 30, 2019, in Pamela Youde Nethersole Eastern Hospital (PYNEH), a 1700-bed regional hospital in Hong Kong. Patients admitted to the ICU in PYNEH with influenza A or pneumococcal pneumonia were enrolled. Patients with co-infection by both viruses and bacteria, age less than 18 years and insufficient data were excluded. Retrospective analysis of medical records, data in clinical management systems and clinical information systems (IntelliVue Clinical Information Portfolio, Philips Medical, Amsterdam, Netherlands) was performed.

The primary outcome was the ability of NLR to predict hospital mortality. Secondary outcomes were the ability of MLR, PLR and RDW to predict hospital mortality and the diagnostic performance of NLR, MLR, PLR and RDW in discriminating viral from bacterial pneumonia. Laboratory parameters included in the study were obtained from the complete blood count (CBC) at 0 h and 48 h of admission. The delta value was obtained by subtracting the data at 0 h from the data at 48 h.

Statistical analysis

Demographics, clinical outcomes, and laboratory parameters were compared between hospital survivors and non-survivors and patients with viral and bacterial pneumonia. Categorical variables are expressed as the number of cases and percentages and continuous variables as the median ± interquartile range (IQR). Univariate analysis for categorical variables was performed using Fisher’s exact test or Pearson’s chi-square test, as appropriate. Continuous variables were compared by the Mann‒Whitney U test or Student’s t test. Variables with P < 0.1 in univariate analysis were included in multivariate analysis. Logistic regression analysis with backwards stepwise elimination was used to assess independent predictors for hospital mortality. A P < 0.05 was considered significant.

Comparison of the prognostic and diagnostic accuracy of variables was carried out using receiver operating characteristic (ROC) curves. The area under the receiver operating characteristic curve (AUROC) was calculated, ranging from 0.5 to 1.0. Higher values show greater power in the discriminatory outcome.

A classification tree model was used to identify predictors for hospital mortality. This data mining method classifies the studied population into subgroups of dependent variables based on values of independent variables by using non-parametric testing. The splitting method is called the exhaustive chi-squared automatic interaction detector (CHAID). The analysis was conducted in a stepwise manner using the Pearson chi-squared test. The variable with the smallest Bonferroni-adjusted p value and yielding the most significant split was chosen. Nodes were created that maximised group differences in the outcome. A terminal node was produced when the number of child nodes was below 2 or when the smallest adjusted p value was insignificant. All analyses were performed using Statistical Package for Social Sciences for Windows, version 27.0 (SPSS, Chicago, United States).

The sample size was calculated based on an average AUROC of 0.688 (average AUROC taken from 3 studies: 0.74623, 0.69524 and 0.62225) for the neutrophil-to-lymphocyte ratio (NLR) in predicting hospital mortality. With a type I error of 0.05, power of 80%, and expected mortality rate of 12.5%, the calculated sample size was 189.

Ethical approval

This study was approved by the Hong Kong East Cluster Ethics Committee of the Hospital Authority (HKECREC-2020-071), which also waived the need for written informed consent due to the retrospective nature of the study. All methods were performed in accordance with relevant guidelines and regulations.


Baseline characteristics and clinical outcomes

A total of 216 patients were enrolled during the 2 years indicated. Their baseline characteristics and clinical outcomes are listed in Table 1. The median age was 69 (interquartile range 59–80). The median APACHE IV score was 91 (63–115). The median APACHE IV score for predicting risk of death was 0.39 (0.16–0.66). Most of the population (83.3%, 180/216) had septic shock. Sixty-nine percent (149/216) of patients were mechanically ventilated, and 34.3% (74/216) required renal replacement therapy. The median intensive care unit (ICU) and hospital length of stay were 4 (1.8–9.7) and 13.3 (7.0–28.8), respectively.

Table 1 Comparison of patient demographics and outcome parameters between viral and bacterial pneumonia.

Univariate analysis (Table 1) showed that diabetes mellitus (43.2% vs. 26.7%, P = 0.015) was more common in patients with viral pneumonia. In comparison, patients with bacterial pneumonia were more likely to develop septic shock (94.3% vs. 73%, P < 0.001).

Comparison between survivors and non-survivors

The overall hospital mortality rate of the enrolled population was 31% (n = 67). Univariate analysis (Table 2) showed that hospital non-survivors were older (median 78 vs. 65, P < 0.001) and more likely to have chronic kidney disease or end-stage renal failure (70.1% vs. 39.6%, P < 0.001) and haematological malignancy (10.4% vs. 3.4%, P = 0.051). Hospital non-survivors also had higher APACHE IV scores (110 vs. 79, P < 0.001), APACHE IV score predicted risk of death (0.61 vs. 0.26, P < 0.001), and these patients were more likely to have septic shock (92.5% vs. 79.2%, P = 0.017), received mechanical ventilation (83.6% vs. 62.4%, P = 0.002), required renal replacement therapy (47.8% vs. 28.2%, P = 0.008), and shorter hospital lengths of stay (11.1 vs. 14.1, P = 0.012).

Table 2 Comparison between hospital survivors and non-survivors.

Univariate analysis of laboratory parameters (Table 2) showed that lymphocytes at 0 h and 48 h and their delta and that monocytes at 48 h and platelets at 0 h and 48 h and their delta were significantly lower in non-survivors. NLR at 48 h and its delta and RDW at 0 h and 48 h were significantly higher in non-survivors.

Prognostic performance of laboratory parameters

The prognostic power of the significant parameters identified in univariate analysis was compared with the APACHE IV score, and APACHE IV predicted risk of death by ROC analysis (Table 3). APACHE IV predicted risk of death (AUC 0.776, 95% CI 0.713–0.84), and the APACHE IV score (AUC 0.769, 95% CI 0.705–0.833) had the highest discriminatory ability for the prediction of hospital mortality. Platelets at 48 h performed the best among the laboratory parameters assessed in predicting hospital mortality (AUC 0.721, 95% CI 0.643–0.798).

Table 3 Area under receiver operating characteristic curves (AUROC) for prediction of hospital mortality.

Variables that were associated (P < 0.1) with hospital mortality in the initial univariate analysis (Table 2) were included in multivariate analysis. Table 4 shows the logistic regression analysis of predictors of hospital mortality. Independent predictors of hospital mortality included age (odds ratio 1.052, P = 0.001), APACHE IV score (OR 1.020, P = 0.001), RDW at 48 h (OR 1.268, P = 0.011), delta NLR (OR 1.019, P = 0.051) and platelet count at 0 h (OR 0.994, P = 0.013). The Hosmer‒Lemeshow test was used to ensure the goodness of fit of statistical models, with a p value of 0.814, which indicated good calibration and model fit.

Table 4 Logistic regression analysis using backwards stepwise regression (likelihood ratio) for independent predictors of hospital mortality.

The composite of the 5 parameters in Table 4 was referred to as Model 1. Figure 1 compares the prognostic performance of Model 1 and the APACHE IV score by ROC analysis. Model 1, being a composite of five independent predictors, showed superiority in predicting hospital mortality (AUC 0.830, 95% CI 0.772–0.888) over the APACHE IV score alone (AUC 0.769, 95% CI 0.705–0.833).

Figure 1
figure 1

Receiver operating characteristic (ROC) curves to compare the performance of the APACHE IV score and Model 1. Model 1 (logistic regression model from Table 4): AUROC 0.830, 95% CI 0.772–0.888. APACHE IV score: AUROC 0.769, 95% CI 0.705–0.833.

Classification tree analysis

The classification tree model (Fig. 2) was applied to analyse determinant factors that predict hospital mortality. The most significant predictor was the APACHE IV score (adjusted P = 0.000, χ2 = 35.591). For patients with APACHE IV scores ≤ 70, those requiring mechanical ventilation had increased hospital mortality (adjusted P = 0.014, χ2 = 5.999, hospital mortality rate of 14% vs. 0% compared with those not requiring mechanical ventilation). For patients with APACHE IV scores > 90 and ages > 78 (adjusted P = 0.007, χ2 = 11.221), the hospital mortality rate reached 67.4%. For patients with APACHE IV score > 90 and age ≤ 78, platelet count (adjusted P = 0.004, χ2 = 12.316) became an important determinant of mortality. Those with platelet counts ≤ 128 at 0 h had higher hospital mortality (59.3% vs. 16.7%) than those patients with platelet counts > 128.

Figure 2
figure 2

Classification tree analysis for predictors of hospital mortality.

Diagnostic performance of laboratory parameters

Patients with viral pneumonia had a lower white cell count (WCC) at 48 h (9.9 vs. 13.6, P < 0.001), neutrophil count at 48 h (8.8 vs. 11.6, P = 0.001), and delta red cell distribution width (RDW, 0.2 vs. 0.3, P = 0.013) than patients with bacterial pneumonia (Table 5). In receiver operating characteristic (ROC) curve analysis of these parameters (Table 6), WCC at 48 h (AUC 0.648; 95% CI 0.572–0.722) had a greater ability to differentiate viral from bacterial pneumonia than neutrophils at 48 h (AUC 0.627, 95% CI 0.552–0.702) and delta RDW (AUC 0.594, 95% CI 0.518–0.670). Figure 3 displays the respective ROC curves of these parameters.

Table 5 Comparison of laboratory parameters between viral and bacterial pneumonia.
Table 6 Area under the receiver operating characteristic curve (AUROC) for diagnostic differentiation of viral versus bacterial pneumonia.
Figure 3
figure 3

Receiver operating characteristic (ROC) curves to compare the diagnostic performance of laboratory parameters in differentiating viral versus bacterial pneumonia.


Neutrophil-to-lymphocyte ratio (NLR) in the prediction of hospital mortality

The NLR has been studied as a marker of severity and prognostication due to its ability to identify states of extreme physiological stress. Its use has been extensive in different diseases and conditions, including rheumatic diseases13,22, acute pulmonary embolism21,26, acute coronary syndrome14,20 and acute pancreatitis27. The use of NLR in predicting the severity of community-acquired pneumonia (CAP) has been intensively studied28. Its performance was shown to be comparable to the pneumonia severity index (PSI)29,30, CURB-6530, WCC and CRP30,31. Previous studies have proven NLR to be a helpful prognostic marker for patients with sepsis12,24,32,33 and, in general, critically ill populations23,34,35. However, there are scarce literature on its use in the prognostication of critically ill CAP patients. To the best of our knowledge, our study is the first to explore the use of NLR in pneumonia patients in the ICU setting. We could not demonstrate a significant difference in NLR between survivors and non-survivors in our critically ill cohort. Therefore, NLR may be a useful screening tool to stratify CAP patients before ICU admission but has limited value in prognostication of the critically ill population.

An interesting observation from our study was the use of delta NLR in the prediction of hospital mortality. We detected a significantly higher delta NLR in non-survivors than in survivors (3.8 vs. 0.0, P = 0.016), which resulted from an elevation in NLR from 0 h (median 10.7, IQR 5.5–17.9) to 48 h (median 15.4, IQR 9.4–31.4). This persistent elevation or lack of improvement in NLR indicated treatment failure over the illness trajectory, making it a marker of poor prognosis. Our findings were consistent with previous studies that had similar observations36,37.

Neutrophil-to-lymphocyte ratio (NLR) in the diagnostic differentiation of pneumonia aetiology

NLR has received significant attention for its diagnostic accuracy in sepsis, pneumonia and bacteraemia7,8,9,10,11,38. Several studies have proven NLR to be at least a moderate predictor of bacteraemia, with AUROCs ranging from 0.7 to 0.778,9,10,11. Compared to other biomarkers, including C-reactive protein (CRP) and procalcitonin (PCT), NLR shows good correlation and comparable performance in diagnosing bacterial sepsis in emergency care settings. In the critically ill population, CRP and PCT appear to be superior to NLR in diagnosing sepsis39,40,41,42. However, limited literature exists on its use to determine underlying microbiological aetiology. Our study investigated the use of NLR in discriminating between viral and bacterial pneumonia and, to our disappointment, was found to be inferior to WCC. Only 2 paediatric studies have investigated NLR in the differentiation of bacterial and viral pneumonia, consistently demonstrating its poor discriminatory power43,44. One possible explanation is that NLR reflects a patient’s physiological stress when critically ill, regardless of microbiological aetiology. To the best of our knowledge, our study is the first to investigate the use of NLR in differentiating between viral and bacterial pneumonia in the adult critically ill population.

Monocyte-to-lymphocyte ratio (MLR)

Monocytes are leukocytes originating from precursors in the bone marrow that are recruited to inflamed tissues via the bloodstream in response to microbial stimuli. Further differentiation into either macrophages or dendritic cells aids effective microbial clearance at infected sites45,46. Mobilisation of monocytes into the peripheral circulation results in an elevated MLR. The MLR has been shown to be useful in the prognostication of rheumatic diseases47, malignancies48, coronary artery diseases49, stroke50 and Guillain‒Barre syndrome51. Recently, its use in different infections has been investigated, including cellulitis52, respiratory virus infection53, pneumonia29,54,55 and bacteraemia56.

The role of MLR as a predictor of clinical outcome has been explored in patients with Klebsiella pneumonia54, correlating positively with mortality and acting as an independent predictor of severe Klebsiella pneumonia, with an AUROC of 0.888 at an optimal MLR cut-off of 0.665. We could not reproduce such a positive correlation between MLR and hospital mortality in our study. The discrepant finding may be explained by the choice of pneumococcus as the representative bacterium in our study, in contrast to Klebsiella, a gram-negative organism. The use of MLR was reviewed by Djordjevic et al.56, who found significantly higher MLR values in patients with gram-negative blood cultures than in those with gram-positive blood cultures.

MLR can also aid in the diagnosis of bacterial and viral infections. Huang et al. reported satisfactory diagnostic performance of MLR in differentiating between patients with community-acquired pneumonia and healthy subjects29. Merekoulias et al. observed monocytosis, lymphopaenia and hence a reduced lymphocyte-to-monocyte ratio (equivalent to a raised MLR) in outpatients infected by the influenza virus during the H1N1 pandemic53. Subsequently, the authors proposed using the lymphocyte-to-monocyte ratio as a screening tool for influenza virus infection, especially at times where the rapid microbiologic test is in great demand.

According to the above studies, MLR may effectively discriminate patients with pneumonia or infected with the influenza virus from healthy subjects. However, its ability to differentiate between the two types of infections is questionable. In our cohort, the monocyte count, lymphocyte count and MLR were not significantly different between the viral and bacterial groups. Hence, MLR did not show significant diagnostic value in distinguishing between viral and bacterial pneumonia. To date, there is no literature on the use of MLR to differentiate different types of pneumonia.

Platelet (PLT) and platelet-to-lymphocyte ratio (PLR)

Platelets are vital in adaptive immunity and in eliciting an inflammatory response in addition to their primary role in haemostasis57. A strong correlation was demonstrated between platelet count and hospital mortality in CAP patients58,59,60. Consistent with previous studies, we showed that non-survivors had significantly lower platelet counts than survivors at both 0 h and 48 h. The predictive performance of platelet count at 48 h (AUROC 0.721) was comparable to the APACHE IV score (AUROC 0.769), with the best performance of all blood parameters in our study.

PLR is increasingly recognised as an indicator of the inflammatory process and has been shown to have good prognostic value in patients with cancers61, acute myocardial infarction15 or stable coronary artery disease62. Its use in prognostication has been extended to the critically ill and septic population, as evidenced by studies showing an association between PLR and ICU length of stay63 and even hospital mortality64,65. Our study was not able to demonstrate such a correlation between PLR and hospital mortality. The difference in the sample size of the cohorts may be a significant factor contributing to the inconsistent findings.

Red cell distribution width (RDW)

The red cell distribution width (RDW) measures variability in red blood cell (RBC) size. Significant associations have also been demonstrated between RDW and patients with sepsis and community-acquired pneumonia66,67,68,69,70,71. Several mechanisms have been proposed to explain the correlation between elevated RDW and inflammatory status. Pro-inflammatory cytokines such as interleukin-1β, interleukin-6, and tumour necrosis factor-α have been shown to shorten RBC survival72. Erythropoietin production and erythroid precursor cell differentiation are suppressed73. Compensatory release of the larger premature RBCs known as reticulocytes into the circulation results in an elevated RDW.

RDW was found to be significantly associated with 30-day mortality when evaluated as a prognostic marker in septic patients at the Emergency Department69,71. Our study was able to demonstrate RDW as an independent predictor of mortality in the critically ill population. Apart from the absolute value of RDW, its change from baseline to 72 h after admission was studied in severely septic patients attending the Emergency Department68, and it was found to be an independent predictor of hospital mortality. Our study evaluated delta RDW, as defined as the change from baseline to 48 h after admission, and we did not reproduce the result of delta RDW as an independent predictor of mortality. However, to our surprise, delta RDW showed marginal diagnostic ability in differentiating between viral and bacterial pneumonia (AUROC 0.594). To our knowledge, there have been no previous studies on the association between RDW levels and the aetiology of CAP.


Our study has several limitations. First, this was a single-centre study with a limited sample size, affecting the generalisability and reliability of the results. Further studies with larger sample sizes may be helpful. Second, this was a retrospective study and potentially confounded by selection bias. Third, we chose Streptococcus pneumoniae as a representative organism for bacterial pneumonia and influenza A for viral pneumonia. Thus, our results may not represent bacterial and viral pneumonia caused by organisms other than pneumococcus and influenza A. Fourth, we did not include novel biomarkers, such as C-reactive protein (CRP) and procalcitonin (PCT), which have been extensively studied and found to be helpful in prognostication for critically ill septic patients. These markers were not readily available in our hospital at the start of our study period and hence were not incorporated to compare the studied biomarkers. Fifth, the treatment provided for pneumonia may have a positive or negative impact on the value of the haematological markers studied, hence affecting evaluation of the diagnostic efficacy of these markers. Last, we did not exclude patients with an immunocompromised state, for instance, patients on long-term corticosteroid use and those infected with human immunodeficiency virus (HIV). These factors can significantly impact the baseline neutrophil count and, hence, the neutrophil-to-lymphocyte ratio, causing confounding in the interpretation of these biomarkers. Additionally, we included patients with active haematological malignancies (5.6%) in our study. The use of NLR has not been validated in this population, and the results should be interpreted with caution.


In predicting the outcome of critically ill CAP patients, the prognostic power of the APACHE IV score is superior to all the blood parameters studied. The addition of other factors that are independent predictors of mortality to the APACHE IV score further strengthens its prognostic power. However, the APACHE IV score is limited by the need for multiple clinical and laboratory parameters, which is deemed less convenient than parameters directly derived from a simple complete blood count. In our cohort, the single, simple biomarker with comparable prognostic performance to the APACHE IV score by the ROC analysis was found to be the platelet count at 48 h. Further studies should be carried out to investigate the use of other novel inflammatory markers, such as CRP and PCT, in critically ill patients with pneumonia. The use of multiple, composite biomarkers, including CRP and PCT, instead of single biomarkers, should also be considered, and their predictive power compared with that of the APACHE IV score.

In determining the aetiology of pneumonia in critically ill patients, no single biomarker has good diagnostic accuracy.