Introduction

Complete blood count (CBC) indices and their ratios, including neutrophil-lymphocyte ratio (NLR) and platelet-lymphocyte ratio (PLR), are widely reported to have significant associations with adverse clinical outcomes in many acutely ill cohorts1,2,3,4,5,6,7. These markers have been studied in diverse clinical settings for their potential to inform diagnosis and prognosis1,2,3,4,5,6,7,8,9, assist with triage10, and guide treatment decisions11,12. The pathophysiologic mechanisms generating these associations in diverse disease settings are not fully understood, and this mechanistic uncertainty may contribute to reduced clinical utilization of these markers. Recent work has identified co-regulation of white blood cell (WBC) and platelet (PLT) populations as a feature of healthy inflammatory recovery13 that is preserved across heterogeneous pathologic processes including different types of trauma, ischemia, and infection. Favorable inflammatory recoveries are defined by exponential WBC decrease followed by linear PLT increase. This finding provides a potential unifying mechanism for CBC ratio associations with mortality across the diversity of acute inflammatory disease settings.

If a universal WBC-PLT inflammatory recovery trajectory helps generate the widely observed associations between CBC index ratios and clinical outcomes, then a ratio specifically designed with this trajectory in mind, PLT/WBC (PWR), should reliably increase during successful inflammatory recoveries and be at least as strongly associated with clinical outcomes as other ratios that do not as directly reflect these WBC-PLT recovery dynamics. Thus, we hypothesized that PWR would provide significant risk stratification in inflammatory settings by reducing the WBC-PLT trajectory to one dimension, with low PWR reflecting failure of PLT recovery, or persistent WBC elevation, or both. However, PWR has not been studied systematically in contrast to NLR, PLR, red blood cell distribution width (RDW), and other CBC-derived statistics.

Here, we show that PWR provides stronger stratification of mortality risk than other CBC markers and ratios in the setting of four different diseases involving acute inflammation: COVID-19, heart failure, myocardial infarction, and stroke. We validate these findings in a cohort from an independent medical center. Our results are consistent with recently identified inflammatory dynamics13 and suggest that PWR may provide a convenient and interpretable summary of patient inflammatory state.

Results

Platelet-white cell ratio is more strongly associated with mortality risk than other blood count markers

Analysis was initially performed on cohorts of patients from Massachusetts General Hospital (MGH) with characteristics in Table 1. The four MGH cohorts represent demographics of patients with these conditions and are heterogeneous, with significant variations in age, gender, self-reported race, blood count markers, and overall mortality. Associations between 30 d all-cause mortality and CBC indices and ratios are shown in Fig. 1A. In all four cohorts, PWR generates the largest hazard ratio for a one standard deviation (std) change. RDW is a widely reported general risk marker14,15,16 and shows the second or third strongest association with mortality in each cohort but is consistently weaker than PWR. NLR was also strongly associated across all four cohorts. Other markers showed significant variation in the strength of mortality associations and relative ranking.

Table 1 Characteristics of the study cohorts
Full size table
Fig. 1: Association of CBC indices and ratios with 30-day mortality.
figure 1

A Age-adjusted 30 d, all-cause mortality hazard ratios for each marker for inpatients with one of four admitting diagnoses at MGH. Hazard ratios correspond to a one standard deviation change in each marker. Red arrows denote markers for which an increase is associated with higher risk, and blue arrows denote markers where a decrease is associated with higher risk. B Kaplan–Meier curves for patients in each cohort stratified by the percentile of the admitting PWR value. The PWR values corresponding to each percentile are listed in each panel. C Mortality hazard ratios for PWR < 10, PWR < 20, and per 5 unit decrease. D PWR-mortality curves calculated via logistic regression. Error bars in (A, C) reflect 95% confidence intervals on the hazard ratio, with cohort n given in Table 1. Results in (A, B) were validated in independent cohorts from the University of Washington Medical Center (Supplementary Fig. 8). Results stratified by gender show similar findings (Supplementary Fig. 12). Abbreviations are defined in the methods. Source data are provided as a Source Data file. Abbreviations: COVID: COVID-19, H.F. heart failure, M.I. myocardial infarction. Marker abbreviations are defined in Table 1.

Full size image

PWR also showed the strongest hazard ratios when markers were standardized by interquartile range instead of standard deviation (Supplementary Fig. 1) and when 7-day mortality or mortality at discharge were considered (Supplementary Fig. 2). Over follow-up periods of 90 days or longer, PWR was still significantly associated with risk, but RDW showed stronger associations for some of the cohorts (Supplementary Fig. 2), consistent with the hypothesis that PWR summarizes acute recovery processes. In a multivariate model, PWR was more strongly associated with mortality than any other CBC marker (Supplementary Fig. 3), and while patient age is the strongest predictor of mortality in many contexts17, for the stroke and heart failure cohorts PWR was surprisingly more strongly associated with short-term mortality than even patient age. PWR also showed stronger mortality associations than PLT/log(WBC) (Supplementary Fig. 4).

Given the consistently large hazard ratios generated by PWR, we performed time-to-event analysis after stratification by PWR (Fig. 1B). Low PWR (<10th percentile) was associated with increased mortality risk compared to moderate PWR (>25th percentile), with increases of 3.7× (2.9–4.7, 5.8%–21.4%) for COVID-19, 2.6× (2.0–3.5, 7.1%–18.6%) for heart failure, 3.7× (3.0–4.7, 5.3%–19.7%) for myocardial infarction, and 3.0x (2.5–3.7, 10.8%–33.0%) for stroke (95% CI, lower risk-higher risk). PWR percentiles (10th/25th) are 16.2/22.7, 13.7/18.6, 13.5/18.2, and 12.0/16.4 for these cohorts respectively. These results were robust to choices of specific threshold values. See Supplementary Fig. 5 for distributions and deciles of PWR for the four cohorts and Supplementary Fig. 6 for Kaplan-Meier curves using quintiles instead of the 10th and 25th percentiles.

All four cohorts showed similar hazard ratios (~1.2–1.3) per 5 unit decrease in PWR, as well as at a fixed thresholds (PWR < 10, and PWR < 20) (Fig. 1C). PWR-mortality curves (estimated via logistic regression) showed similar shape across the four cohorts, with the most rapid increases in mortality occurring near a PWR value of 25 (Fig. 1D).

Analysis of the COVID-19 cohort after stratification by ‘wave’ (a proxy for dominance of different COVID-19 variants) showed that PWR remained a strong predictor over all waves (Supplementary Fig. 7). For wave 1 (Mar–Oct 2020) and wave 2 (Oct 2020–July 2021) PWR was the strongest predictor, while in wave 3 (July 2021–Apr 2022) it was the second strongest behind ALYMPH. We also note that ALYMPH was highly variable as a predictor, having a statistically insignificant hazard ratio for wave 2, and a substantially lower ratio than PWR for wave 1.

Comparison to disease-specific prognostic markers

PWR appears to have prognostic value superior to other non-specific blood count markers in the contexts of these four diseases, but in some cases, disease-specific prognostic biomarkers are available: high-sensitivity Troponin-T (HS-TnT) for myocardial infarction18, N-terminal pro-brain natriueretic peptide (NT-proBNP) in heart failure19,20, and D-dimer has been studied in COVID-1921 but is less validated. We therefore investigated whether PWR might complement these disease-specific markers in the appropriate cohorts (Fig. 2). There is no well-established and widely collected disease-specific biomarker for stroke22. NT-proBNP has been proposed by some investigators23 but was not available for enough subjects in our cohort for a comparison to be informative. In all three cohorts, PWR was significant in multivariate models, suggesting PWR captures complementary signals of risk, and CBCs are often more accessible in general24 and were in our cohorts (Fig. 2B).

Fig. 2: PWR complements disease-specific prognostic markers.
figure 2

A Univariate and multivariate mortality hazard ratios were calculated for PWR and disease-specific prognostic markers in the MGH cohort: D-dimer for COVID-19, NT-proBNP for heart failure, and HS-TnT for myocardial infarction. Disease-specific biomarkers for heart failure and myocardial infarction generated higher univariate hazard ratios than PWR. For all three cohorts, PWR provided statistically significant prognostic information in multivariate modeling suggesting it provided complementary information. B Percentage of patients for whom each disease-specific marker was available within 48 h of admission. Error bars (A) reflect 95% confidence intervals on the hazard ratios, with each cohort n given in Table 1. For HS-TnT data availability was calculated only after Apr-2018, when MGH switched to use of a high-sensitivity troponin assay. PWR hazard ratios may differ slightly from those in Fig. 1, as they are calculated only in the subset of patients who have the alternative marker available. In (A) univariate refers to hazard ratios that are corrected for age, but not for the other marker, and multivariate refers to jointly fitting both markers along with age. D-dimer, NT-proBNP and HS-TnT have long-tailed distributions and were log-scaled before analysis. Source data are provided as a Source Data file. Marker abbreviations are defined in Table 1.

Full size image

Validation studies at an independent medical center

To test the generalizability and stability of findings in Fig. 1, we replicated analysis in four similar cohorts at the University of Washington Medical Center (UWMC, Supplementary Table 1). PWR in the UWMC cohorts consistently produced the highest mortality hazard ratios among all the blood count markers (Supplementary Fig. 8). RDW was also one of the top 3 markers at both sites, except for in the UWMC COVID-19 cohort. Relative rankings of other markers were less consistent between centers. With the exception of COVID-19, the MGH and UWMC cohorts have similar PWR distributions and hazard ratios (Fig. 3). The results for COVID-19 may reflect timing-related differences in pathogen strain and clinical management protocols. Similarly, the PWR sensitivity for mortality prediction with a 75% specificity was similar at the two sites (54% v 47% for COVID-19, 40% v 40% for heart failure, 48% v 49% for myocardial infarction, and 46% v 51% for stroke; ordered as MGH and UWMC). This consistency suggests that while calibration of PWR thresholds may need to be performed for different disease cohorts, it may be less necessary between medical centers, though further validation is necessary.

Fig. 3: Validation of results in independent cohorts at the University of Washington Medical Center (UWMC).
figure 3

A PWR mortality hazard ratios for cohorts at MGH and UWMC were consistent. B Distributions of admitting PWR values at MGH and UWMC were also similar. Error bars (A) reflect 95% confidence intervals on the hazard ratios, with each cohort n given in Table 1. Note that due to differences in data availability at the two medical centers, the COVID-19 cohorts have significantly different median admit dates (MGH: Apr-15-2021, UWMC: Jan-01-2022) and thus contain substantially different mixes of disease variants. Source data are provided as a Source Data file. Marker abbreviations are defined in Table 1.

Full size image

Validation in a pre-operative general surgical cohort

To test the generalizability of these results beyond the four diseases considered, we evaluated PWR in a more general and less well-controlled cohort available in the MOVER database25, a publicly available dataset of surgical admissions and outcomes from University of California Irvine Medical Center (UCIMC). This dataset comprises a wide array of surgeries, ranging from relatively benign (e.g., excision of a skin lesion) to more severe (e.g., a mitral valve replacement), all of which are expected to be associated with sterile trauma and an associated inflammatory response. For this evaluation, we used the first available CBC + differential in the 48 h before each patient’s first surgery in the database. Many patients with elective procedures may be at baseline at the time of this CBC, and for those individuals, PWR reflects baseline inflammatory state. For other patients whose surgery was indicated because of an acute pathologic condition, the pre-operative PWR likely reflects the inflammatory response to the stimulus that preceded surgery. PWR in this far more heterogeneous and unselected cohort still demonstrated the strongest 30-day mortality hazard ratio of any CBC marker or ratio (Fig. 4A), with strong separation of Kaplan-Meier curves by PWR percentile (Fig. 4B). PWR may thus provide prognostic value in less specific patient triage settings even when information only on baseline inflammatory state is available.

Fig. 4: Association of CBC markers and ratios with mortality in a general surgical cohort.
figure 4

A Associations of pre-operative marker values (within 48 h of surgery start) with discharge mortality in a general surgical cohort from the University of California Irvine Medical Center. B Kaplan–Meier analysis by PWR value is similar to Fig. 1 in this general surgical cohort. Error bars (A) reflect 95% confidence intervals on the hazard ratios (n: 9351 patients). Source data are provided as a Source Data file. Marker abbreviations are defined in Table 1.

Full size image

Discussion

The complete blood count is a cornerstone of modern clinical evaluation24 and has been shown in many studies to have surprising and unexplained associations with adverse outcomes particularly in acute inflammatory states1,2,3,4,5,6,14,26,27, but this prognostic information is often unfamiliar to physicians and represents unrealized clinical value. Based on analysis of patient cohorts at three different clinical sites, this study identifies PWR as a convenient summary of patient inflammatory state in rapidly changing acute care settings.

Infection, ischemia, and trauma all trigger inflammatory responses that mobilize WBC and consume PLT with feedback to WBC and PLT proliferation rates. Healthy individuals mount an effective response, and their WBC and PLT populations tend to recover along a universal WBC-PLT trajectory13. PWR provides a single marker encapsulation of these dynamics, reducing this WBC-PLT trajectory to one dimension, with abnormally low PWR reflecting failure of PLT recovery, persistent WBC elevation, or both. Low PWR may reflect dysregulation between PLT and WBC populations analogous to the way elevated RDW can reflect perturbations to red cell homeostasis and provide prognostic information on short-term risk14,15,16,27,28. The consistency of these changes across heterogeneous acute inflammatory settings13 suggests that PWR may succinctly quantify key aspects of inflammatory regulation and recovery, and that PWR changes may occur as part of, or in tandem with a normal inflammatory response and recovery. Other non-inflammatory processes that alter WBC or PLT will also influence PWR and may limit its prognostic utility in inflammatory states. The consistency of PWR’s stronger association with short-term mortality compared to other CBC-derived markers in 4 heterogenous study cohorts across two medical centers and in a general surgical cohort from a third independent medical center supports the prior finding of universal WBC-PLT dynamics during recovery from acute inflammatory stimuli13. This mechanistic explanation29,30 for the associated prognostic signal may facilitate clinical utilization.

In two settings where accepted disease-specific prognostic markers are available, HS-TnT in MI and NT-proBNP in heart failure, PWR appears to provide complementary risk information. HS-TnT is a direct measure of heart tissue damage18, and NT-proBNP elevation is driven by increased myocardial wall stress19,20. PWR may thus be providing nonspecific information on hematologic aspects of the inflammatory response that complement  these tissue-specific signals (Supplementary Fig. 9). PWR is based on CBC indices which are collected frequently in most major clinical settings24 and can provide some prognostic information when these other disease-specific markers are not available.

There was more concordance between medical centers than between diseases at each site (Fig. 3), and calibration of PWR for use with particular disease cohorts is important, including comparison to a general healthy population31,32 to establish specificity (Supplementary Fig. 10), and reference populations may be need to be conditioned on demographics, physiology, as well as current disease state13,33,34,35. Nevertheless, blood count markers and ratios are already sometimes used in the non-specific prognostic assessment of patients, and these results suggest PWR may provide more accurate prognostic information in at least some of those settings3,5,6,7,26.

This study is limited by its retrospective and observational nature. While results were validated across multiple medical centers, there is always potential for unmeasured confounding. Results were age-corrected, but did not control for patient health status (e.g., comorbidities, etc.) beyond the primary cohort diagnosis. We also note that the UCIMC database  includes  discharge mortality instead of the 30-day post-discharge mortality used in the MGB and UWMC analyses. Given these limitations, prospective validation is necessary, and this report also motivates consideration of studies that use PWR to help guide prognostic evaluation, patient triage, and clinical decision making. Prior findings identify deviations from a universal inflammatory recovery trajectory13 as a contributor to PWR variation, but since much of the PWR heterogeneity analyzed in this study was present prior to initiation of an inflammatory recovery, it will be important for future investigations to analyze the initial phase of the inflammatory response. By building on this developing understanding of recovery phase dynamics and PWR, it should be possible to derive significant insights into the circulating cell population mechanisms generating this observed variation in both PWR and associated clinical risk.

Methods

Our study was performed under research protocols approved by the institutional review boards at Massachusetts General Hospital (for data analysis of Mass General Brigham cohorts), and at the University of Washington (for data analysis of University of Washington Medical Center cohorts). For the UWMC and MGB cohorts, the study involved no patient interaction, and each IRB approved a waiver of consent and temporary access to patient identifiers for the duration of the study. Data from the MOVER cohort was de-identified prior to access and was performed in accordance with a research data use agreement. There was no direct interaction with any patients at any point during this study.

Study design

The study’s objective was to compare mortality associations for blood count markers and ratios and to determine whether any marker consistently showed stronger associations in the setting of different diseases. Based on prior research13 we hypothesized that a low PWR ratio would show a stronger mortality associations than other blood count markers and ratios.

Analysis was first performed in four inpatient cohorts from MGH. Choice of the four cohorts was made a priori to reflect four common scenarios involving different acute inflammatory perturbations (infection, ischemia, organ failure). Data collection start points were chosen based on data availability within the electronic health record system. End points were chosen based on the date of initial data collection. Inclusion and exclusion criteria (detailed below) as well as primary endpoints were chosen a priori to reflect typical acute admissions (i.e. >24 h hospital stay), and based on data availability constraints (needing at least one blood count early in admission). No outliers were excluded from analysis. The study design for selection of markers and ratios detailed below was decided a priori. Methods for normalization and comparison of hazard ratios were decided a priori, and investigations of robustness to these assumptions are explored in supplementary materials.

Key findings were validated using four similar cohorts derived from UWMC. A secondary validation was performed using the publicly available MOVER surgical database from UCIMC25. For both validations, data was collated and analyzed only after the primary MGH results were generated and analyzed.

This study was observational and retrospective relying on electronic health data collected as part of standard clinical care practices, and randomization and blinding were not relevant. Demographic characteristics of the study cohorts are given in Table 1. Where appropriate key results were regenerated after stratification by patient self-reported gender.

Data collection

MGH cohorts were chosen to reflect diverse, high-burden settings of acute infection, ischemia, and decompensation. Each cohort comprised all adult admissions to MGH with the associated primary admitting diagnosis (COVID-19 infection, heart failure, myocardial infarction, stroke; based on ICD10 codes) between Jan-2016 and June-2022 (Mar-2020-June-2022 for COVID-19). Patients with <24 h hospital stays or without a CBC within 48 h of admission were excluded. For each patient, the CBC closest to 1 h post admission was analyzed. 30-day post-discharge, all-cause mortality was the primary outcome. CBCs were only considered if they included a white cell differential. The majority of patients had a CBC within 3 h of admission, and no significant associations were seen between time of first marker measurement and marker value (Supplementary Fig. 11).

For UWMC, each cohort comprised all adult admissions with the same associated primary admitting diagnosis as above between Jan-2010 and Dec-2023 (Mar-2020 and Dec-2023 for COVID-19). Due to a shift in UWMC’s primary EHR system in Apr-2021, most (~60%) patient admissions with sufficient data availability occurred after this date. For each cohort, the same restrictions on CBC availability and length of stay were applied as in the MGH cohort, and the same primary outcome was used.

For the UCIMC validation, we considered all surgeries in the database, which spanned a four-year period. To test consistency of findings in a different context, marker associations with mortality were evaluated pre-operatively. Analysis was limited to adult patients with at least 1 CBC + differential in the 48 h pre-surgery. If multiple CBCs were available, the closest one to 48 h pre-surgery was taken. The primary outcome was mortality at discharge.

Data from MGH and UW were collected from their respective electronic data warehouses, under separate research protocols approved by the corresponding local institutional review boards. UCIMC data is publicly available25 and was used solely for research purposes.

Complete blood count index and biomarker selection

A literature search was conducted to determine which CBC ratios were most frequently studied and/or used. CBC index ratios were selected for inclusion if they appeared in >2 of the top 100 Google Scholar results (ordered by ‘relevance’) for “blood count” + “ratio”. From this literature search, four blood count ratios were identified: NLR (68/100 articles), PLR (27/100), monocyte-lymphocyte ratio (MLR, 7/100), and red cell distribution width (RDW)-platelet ratio (RPR, 3/100). Motivated by the recently reported universal WBC-PLT inflammatory recovery trajectory13, PWR was also included, despite appearing in 0/100 published articles (excluding our own recent study on WBC-PLT dynamics13). All CBC indices that were used in any of the above ratios were also analyzed: PLT, WBC, RDW, and absolute lymphocyte (ALYMPH), monocyte (AMONO), and neutrophil counts (ANEUT).

Statistical analysis

For each index or ratio, mortality associations were calculated by fitting a Cox proportional-hazards model, adjusting for patient age as a covariate. Key results were also analyzed after stratifying by gender in Supplementary Fig. 12. Models were fit using continuous values without thresholding. Patients who survived past 30 days post-discharge were censored at this point. No surviving patients required censoring prior to this point. To allow fair comparison given differences in marker magnitude and variance, each mortality hazard ratio was then normalized based on the marker’s standard deviation. Each hazard ratio presented therefore reflects the increased risk from a one standard deviation change in the marker. To facilitate comparison, ratios below 1 were inverted. Based on standard practice36, the proportional hazards assumption was evaluated by visual inspection of the scaled Schoenfeld residuals37 which revealed no obvious patterns. Hazard ratios reflect the average hazard over each covariate, regardless of proportionality, because none of the surviving patients required early censoring36. Kaplan–Meier estimators were also used to perform time-to-event analysis in the cohort subgroups stratified by marker values. To allow for consistent group sizes across the four cohorts, each cohort was split based on marker percentiles (<10th, 10–25th, and >25th). These splits were chosen to allow meaningfully large numbers of patients in each subgroup. All analysis was performed in MATLAB 2021b and 2023b.

Reporting summary

Further information on research design is available in the Nature Portfolio Reporting Summary linked to this article.