Developing and validating a prognostic model to predict ICU mortality in patients with sepsis-associated thrombocytopenia: a retrospective cohort study based on MIMIC-IV

Abstract

Objective

Given the high morbidity and mortality of patients with sepsis-associated thrombocytopenia (SATP) in the intensive care unit, this study retrospectively analysed the influencing factors for poor prognosis in patients with SATP using the MIMIC database, constructed a nomogram model and verified the predictive performance of the model.

Design

A retrospective cohort study.

Setting

The data from MIMIC-IV, V.2.2.

Clinical characteristics

The clinical features of SATP, including demographics, comorbidities, vital signs, laboratory parameters, treatments and clinical management, were extracted from the MIMIC-IV database.

Methods

1409 patients with SATP were included in this study and randomly divided into a training set and a validation set in a ratio of 7:3. The least absolute shrinkage and selection operator and multivariable Cox regression analysis were used to determine the optimal predictors and establish a prediction model. The receiver operating characteristic curve, calibration curve and decision curve analysis (DCA) were used to verify the accuracy and application value of the model.

Results

The nomogram model incorporates nine factors, including clinical characteristics, laboratory test indicators, comorbidities and treatment methods, which were identified as predictors of SATP and used to construct the model. The area under the curve of the model was 0.868 (95% CI: 0.794 to 0.942) in the training set and 0.836 (95% CI: 0.681 to 0.991) in the validation set. The calibration curve and DCA confirmed the clinical application value of the nomogram.

Conclusions

The constructed nomogram for predicting patients with SATP has favourable predictive ability and is helpful to further optimise clinical management strategies.

STRENGTHS AND LIMITATIONS OF THIS STUDY.

• Monitoring the associated risk factors and accurately assessing patient prognosis early, as guided by our model, are crucial and allows us to implement appropriate interventions, thereby improving the prognosis of patients with sepsis-associated thrombocytopenia (SATP).

• It focuses specifically on patients in the intensive care unit and is applicable to all individuals with SATP. The broader inclusion criteria for adult patients with SATP enhance the predictive model’s generalisability. This study incorporated a wide array of independent risk factors, including vital signs, laboratory results, comorbidities and treatment measures, all readily obtainable soon after admission, ensuring the model’s simplicity and timely application.

• Furthermore, the area under the curve and calibration curves demonstrated that the nomogram exhibited favourable discriminative ability and calibration, suggesting substantial potential for clinical application.

• The data were extracted from the MIMIC-IV database; despite the nomogram model being validated with the validation cohort, further validation using multicentre data is required.

• The model is not applicable to patients with missing variables; there may be variables not included in the model that could also influence in-hospital mortality in patients with SATP.

Introduction

The Sepsis-3 diagnostic criteria define sepsis as a condition marked by organ dysfunction, which can be attributed to a dysregulated host response to infection. Meanwhile, septic shock represents a more severe manifestation, where the aforementioned sepsis is accompanied by significant disturbances in circulation, cell function and metabolic processes.¹ Despite the publication of sepsis treatment guidelines and advancements in antibiotic therapy and organ support technologies, the mortality rate for patients with sepsis remains as high as 25% or higher.² This severe infectious state triggers the activation of the body’s immune and coagulation systems, precipitating a series of pathophysiological changes. Among these changes, thrombocytopenia frequently manifests in patients with sepsis, posing significant challenges to treatment and jeopardising patient safety. Thrombocytopenia is a prevalent haematologic abnormality observed in more than 60% of patients in the intensive care unit (ICU).³ It can influence short-term and long-term outcomes of these patients, such as heightened haemorrhage risk, increased need for transfusions, utilisation of life support, prolonged ICU and hospital stays and mortality.⁴ Prior research has demonstrated that platelets are of critical importance in the development of sepsis,^{5 6} and thrombocytopenia occurring 24 hours prior to septic shock onset increases the risk of significant haemorrhage, kidney injury and prolonged hospitalisation.⁷ Thrombocytopenia in patients with sepsis develops through diverse mechanisms. In sepsis, platelets become activated and adhere to endothelial cells, resulting in platelet sequestration and destruction.⁸ Concurrently, thrombocytopenia can contribute to dysregulated inflammatory and immune responses in the host, thereby increasing mortality among patients with sepsis.⁹ Thrombocytopenia represents a critical complication of sepsis that significantly impacts patient prognosis.

While several studies have explored the influence of thrombocytopenia on the prognosis of patients with sepsis, there is currently no established predictive model for reliably forecasting the outcomes of patients with sepsis-associated thrombocytopenia (SATP). The development of a predictive model for the prognosis of patients with SATP can assist clinicians in devising personalised treatment approaches, ultimately enhancing patient outcomes. Therefore, our study analysed predictors for in-hospital mortality among patients with SATP by using data from the MIMIC-IV database. Then, we employed least absolute shrinkage and selection operator (LASSO) regression to select potential predictive factors and used multivariable Cox regression to develop a prognostic model to predict the risk of death in ICU patients with SATP, aiming to refine treatment strategies and improve patient survival rates.

Data and methods

Data source

Structured Query Language with pgAdmin4 was employed to query and extract data from the MIMIC-IV, V.2.2 database.¹⁰ The MIMIC-IV 2.2 database is a large public database that contains hospitalisation information for patients at the Beth Israel Deaconess Medical Center between 2008 and 2019. Because the present study was an analysis of a third-party anonymised publicly available database with pre-existing institutional review board (IRB) approval, our institution’s IRB approval was exempted. In the database, the true identity information about the patient is hidden. Thus, obtaining the patient’s informed consent was not required. The author (WY) completed the relevant course training and obtained the certificate (No: 62177075) to access the database. All data are from the PhysioNet official website (https://mimic.physionet.org/).

Patient and public involvement

Patients and/or the public were not directly involved in this study.

Patients

Patients diagnosed with sepsis using the International Classification of Diseases, 9th and 10th editions (ICD-9 and ICD-10 codes: ‘99592’, ‘A419’, ‘R6521’, ‘R6520’, ‘A4159’ and ‘A4150’), and concurrently diagnosed with thrombocytopenia (ICD codes: ‘2875’, ‘D696’, ‘D6959’, ‘28749’, ‘2874’ and ‘2841’) on their initial ICU admission were included in the study. Patient age was limited to 18–100 years. Those discharged from the ICU within 24 hours were excluded. For those with multiple ICU admissions, only their initial ICU admission record was retained for analysis. From the MIMIC database, 1517 adults diagnosed with SATP were selected. Patients with anomalous data (eg, negative ICU stay times) (n=10) and those who stayed in the ICU for less than 24 hours (n=98) were excluded. Ultimately, 1409 eligible subjects were included in our dataset. The entire dataset was split into a training set (n=986) and a test set (n=423) in a proportion of 7:3. Clinical and demographic features did not significantly differ between the training and validation sets (online supplemental table 1).

Data extraction

Patient data within the first 24 hours postadmission were extracted from the MIMIC-IV database. The extracted data encompassed: (1) demographics, including gender and age; (2) comorbidities, including hypertension, type 2 diabetes mellitus, heart failure, chronic kidney disease and pneumonia; (3) vital signs, including body temperature, heart rate, oxygen saturation (SPO₂), respiratory rate and mean arterial pressure; (4) laboratory parameters, including complete blood count: white blood cell count, red blood cell count, platelet count, haemoglobin and red cell distribution width (RDW); comprehensive metabolic panel: total bilirubin, anion gap, chloride, potassium, sodium, calcium, glucose, creatinine and blood urea nitrogen (BUN); blood gas: pH, lactate, partial pressure of oxygen (PaO₂) and partial pressure of carbon dioxide (PaCO₂); coagulation function: prothrombin time (PT), partial thromboplastin time (PTT) and international normalised ratio (INR) and (5) treatments and clinical management, including mechanical ventilation, renal replacement therapy, corticosteroid use and vasoactive drug use, as well as Sequential Organ Failure Assessment (SOFA) score and Simplified Acute Physiology Score II (SAPS-II). To optimise statistical power, all eligible patients from the MIMIC-IV database were included. Missing data bias was minimised by excluding variables with more than 20% missing values from the final cohort. When the proportion of missing values for a variable was less than 20%, multiple imputation was adopted to fill in the missing portions of the variable.

Statistical analysis

Descriptive statistics included both continuous and categorical variables. Comparisons of baseline characteristics were made between survivors and non-survivors in the training cohort. Data that did not adhere to the normal distribution were expressed as the median and IQR and were compared using the Mann-Whitney U test. In contrast, data that exhibited a normal distribution were expressed as mean and SD, with the Student’s t-test employed for comparison. Frequencies and percentages were used to express categorical variables, while the χ² test was applied to evaluate group comparisons. For model development, we applied LASSO-penalised Cox proportional hazards regression for variable selection. We used 10-fold cross-validation to perform variable shrinkage and selection from the 29 variables. The number of variables included in the final model was determined based on the position corresponding to λ.1se. Based on the variables selected from the LASSO-Cox regression, the final model coefficients were derived from standard multivariable Cox regression using the selected variables to ensure HR interpretability. The final Cox proportional hazards model coefficients were then used to construct a prognostic nomogram for predicting the 90-day mortality risk. The accuracy of the nomogram was assessed using the receiver operating characteristic (ROC) curve and quantified by the area under the curve (AUC). Calibration of the nomogram’s predicted probabilities against observed outcomes was evaluated using calibration curves. Decision curve analysis (DCA) was employed to evaluate the clinical utility of our predictive model. To further validate the model, the included patients were divided into two groups (low-risk group and high-risk group), and Kaplan-Meier curves were plotted to compare survival outcomes between the two groups. In order to ascertain whether the findings were statistically significant, two-sided statistical tests were employed, with a p value of less than 0.05 signifying statistical significance. Subsequently, R (V.4.2.0) was used for all the statistical analyses.

Results

Clinical features

Table 1 provides a thorough comparison of the clinical characteristics between survivors and non-survivors in the training group, including information on age, vital signs, comorbidities, interventions and laboratory results. In the training group, 308 patients (31.2%) were ICU non-survivors, and 678 patients (68.8%) were survivors. We found that the occurrence of pneumonia (43.66% vs 55.19%, p<0.001) was significantly associated with mortality. In terms of vital signs, an increased heart rate (91.36 vs 94.74 bpm, p=0.003), lower mean arterial pressure (73.01 vs 71.26 mm Hg, p=0.007) and lower SPO₂ (96.71% vs 96.17%, p=0.015) were associated with mortality. Compared with the survivor group, the mortality group showed a decrease in red blood cell count (3.35 vs 3.17×10¹²/L, p<0.001) and haemoglobin levels (102.3 vs 96.9 g/L, p<0.001), while RDW was significantly increased (15.93% vs 17.08%, p<0.001). Serum potassium levels were increased (4.13 vs 4.33 mmol/L, p<0.001), serum calcium levels were increased (7.85 vs 8.07 mmol/L, p<0.001) and serum chloride levels were decreased (104.94 vs 103.82 mmol/L, p=0.037). The anion gap was increased (15.79 vs 17.31 mmol/L, p<0.001), BUN levels were increased (37.44 vs 45.92 mg/dL, p<0.001) and total bilirubin levels were increased (2.68 vs 5.96 mg/dL, p<0.001). In addition, pH levels were decreased (7.35 vs 7.33, p=0.010), lactate levels were increased (2.50 vs 3.32 mmol/L, p<0.001), PT was increased (18.78 vs 21.39 s, p<0.001), PTT was increased (40.53 vs 46.04 s, p<0.001) and INR was increased (1.75 vs 1.98, p=0.002). Moreover, a higher proportion of non-survivors required vasopressor support (71.68% vs 81.82%, p<0.001) and mechanical ventilation support (44.99% vs 72.40%, p<0.001). Non-survivors also had significantly higher SOFA scores (8.69 vs 10.14, p<0.001) and SAPS-II scores (44.92 vs 52.46, p<0.001) compared with survivors.

Table 1. Baseline characteristics of included patients.

Variables	Total (n=594)	Survivors (n=462)	Non-survivors (n=132)	P value
Age, median (IQR)	67.00 (56.00 to 78.00)	66.00 (55.00 to 78.00)	68.00 (57.00 to 78.00)	0.212
Gender, n (%)				0.620
Male	578 (58.62)	401 (59.14)	177 (57.47)
Female	408 (41.38)	277 (40.86)	131 (42.53)
Laboratory test results
White blood cells, ×109/L, median (IQR)	12.86 (8.01 to 18.70)	12.96 (8.04 to 18.40)	12.58 (7.90 to 19.76)	0.997
Red blood cells, ×1012/L, median (IQR)	3.22 (2.78 to 3.74)	3.31 (2.84 to 3.78)	3.06 (2.67 to 3.55)	<0.001
Platelet count, ×109/L, median (IQR)	108.00 (71.54 to 153.16)	110.75 (75.12 to 152.83)	99.63 (66.14 to 153.66)	0.141
Haemoglobin, g/L, median (IQR)	98.5 (85.2 to 112.5)	100.7 (86.7 to 115.0)	92.1 (83.0 to 108.8)	<0.001
RDW, %, median (IQR)	15.70 (14.35 to 17.43)	15.40 (14.10 to 17.00)	16.53 (14.78 to 18.64)	<0.001
Sodium, mmol/L, median (IQR)	138.00 (135.00 to 141.00)	138.00 (135.00 to 141.00)	138.00 (134.19 to 141.33)	0.591
Potassium, mmol/L, median (IQR)	4.10 (3.75 to 4.55)	4.05 (3.71 to 4.50)	4.21 (3.83 to 4.77)	<0.001
Calcium, mmol/L, mean (SD)	7.92±0.78	7.85±0.76	8.07±0.82	<0.001
Chloride, mmol/L, mean (SD)	104.69 (100.00 to 109.00)	105.00 (101.25 to 108.95)	103.88 (98.31 to 109.25)	0.016
Glucose, mg/dL, median (IQR)	131.38 (107.70 to 169.93)	131.10 (107.80 to 166.97)	132.42 (107.00 to 173.98)	0.768
Anion gap, mEq/L, median (IQR)	15.67 (13.41 to 18.33)	15.33 (13.00 to 17.50)	16.71 (14.24 to 19.52)	<0.001
Total bilirubin, mg/dL, median (IQR)	1.30 (0.70 to 2.84)	1.30 (0.70 to 2.40)	1.30 (0.80 to 3.95)	0.003
Urea nitrogen, mg/dL, median (IQR)	33.47 (19.69 to 52.95)	31.29 (18.50 to 50.00)	39.53 (24.00 to 59.00)	<0.001
Creatinine, mg/dL, median (IQR)	1.47 (0.98 to 2.42)	1.40 (0.95 to 2.33)	1.67 (1.08 to 2.66)	0.003
pH, mean (SD)	7.34±0.07	7.35±0.07	7.33±0.08	0.010
PaCO2, mm Hg, median (IQR)	38.60 (34.75 to 42.50)	38.60 (35.00 to 42.00)	38.60 (34.55 to 43.33)	0.596
PaO2, mm Hg, median (IQR)	88.83 (67.89 to 117.95)	88.83 (67.54 to 114.41)	88.84 (68.83 to 128.85)	0.183
Lactate, mmol/L, median (IQR)	2.26 (1.65 to 3.27)	2.25 (1.55 to 2.96)	2.62 (1.84 to 4.21)	<0.001
PT, s, median (IQR)	16.50 (14.30 to 21.34)	16.50 (14.18 to 19.70)	18.16 (14.76 to 23.92)	<0.001
PTT, s, median (IQR)	36.90 (30.96 to 47.05)	36.90 (30.63 to 43.27)	39.65 (32.92 to 53.64)	<0.001
INR, median (IQR)	1.50 (1.30 to 2.00)	1.50 (1.30 to 1.80)	1.68 (1.35 to 2.25)	<0.001
Vital signs
Heart rate, bpm, mean (SD)	92.41±16.79	91.36±16.81	94.74±16.54	0.003
Non-invasive mean blood pressure, mm Hg, median (IQR)	71.96 (66.48 to 77.26)	72.12 (67.38 to 77.97)	71.24 (64.20 to 75.81)	<0.001
Respiration rate, bpm, mean (SD)	21.45±5.53	21.34±4.43	21.67±7.39	0.385
SPO2, %, mean (SD)	96.78 (95.50 to 98.07)	96.85 (95.69 to 98.09)	96.66 (95.07 to 98.03)	0.032
Temperature, °C, mean (SD)	36.81±1.69	36.84±1.84	36.74±1.29	0.363
Severity scores
SOFA, mean (SD)	9.14±3.89	8.69±3.62	10.14±4.26	<0.001
SAPS-II, median (IQR)	46.00 (36.00 to 57.00)	44.00 (34.25 to 53.00)	51.50 (41.00 to 61.00)	<0.001
Comorbidities
Hypertension, n (%)	341 (34.58)	236 (34.81)	105 (34.09)	0.826
Diabetes, n (%)	281 (28.50)	185 (27.29)	96 (31.17)	0.211
Heart failure, n (%)	297 (30.12)	192 (28.32)	105 (34.09)	0.067
Chronic kidney disease, n (%)	210 (21.30)	134 (19.76)	76 (24.68)	0.081
Pneumonia, n (%)	466 (47.26)	296 (43.66)	170 (55.19)	<0.001
Medical treatment
Ventilation, n (%)	528 (53.55)	305 (44.99)	223 (72.40)	<0.001
Vasoactive, n (%)	738 (74.85)	486 (71.68)	252 (81.82)	<0.001

Data are presented as n (%) for categorical variables, mean (SD) for normally distributed continuous variables and median (IQR) for non-normally distributed continuous variables. Comparisons between survivors and non-survivors were performed using Student’s t-tests (normal)/Mann-Whitney U tests (non-normal) for continuous variables.

INR, international normalised ratio; PaCO₂, partial pressure of carbon dioxide in artery; PaO₂, partial pressure of oxygen; PT, prothrombin time; PTT, partial thromboplastin time; RDW, red cell distribution width; SAPS-II, Simplified Acute Physiology Score II; SOFA, Sequential Organ Failure Assessment; SPO₂, oxygen saturation.

Selection of predictive factors and model construction

Before LASSO regression, variables with a variance inflation factor greater than 5 were excluded (online supplemental table 2). A total of 10 variables—age, platelet count, RDW, total bilirubin, BUN, lactate, SPO₂, PTT, mechanical ventilation use and pneumonia—were included in the multivariable regression. The paths of variable shrinkage and cross-validation results are shown in figure 1A,B. Through multivariable Cox regression analysis, the following nine statistically significant variables were selected for the model. Among these, age, RDW, total bilirubin, BUN, lactate, pneumonia and mechanical ventilation were identified as risk factors, while platelet count and SPO₂ were identified as protective factors. There is no multicollinearity among all potential predictor variables (online supplemental table 3). Patients with pneumonia had a 1.3 times higher risk of mortality (HR: 1.278, 95% CI: 1.012 to 1.613). Each unit increase in lactate was associated with a 1.14 times higher risk of mortality (HR: 1.142, 95% CI: 1.083 to 1.204). Patients requiring mechanical ventilation had a 2.2 times higher risk of mortality compared with those who did not (HR: 2.2, 95% CI: 1.687 to 2.871). For age (HR: 1.015, 95% CI: 1.007 to 1.023), RDW (HR: 1.047, 95% CI: 1.005 to 1.090), total bilirubin (HR: 1.029, 95% CI: 1.015 to 1.044) and BUN (HR: 1.004, 95% CI: 1.000 to 1.008) each unit increase was associated with a minor increase in patient mortality risk. Conversely, each unit increase in platelet count was associated with a 0.998 times lower risk of mortality (HR: 0.998, 95% CI: 0.996 to 1.000), and each unit increase in SPO₂ was associated with a 0.934 times lower risk of mortality (HR: 0.934, 95% CI: 0.909 to 0.959) (table 2). A nomogram was constructed using these variables to predict the 90-day mortality rate (figure 2). Scores for each prognostic factor were obtained by finding the matching values on the vertical line. The nomogram’s total score was determined by adding the results of each prognostic factor. A vertical line drawn from the total score axis was used to calculate the estimated 90-day mortality risk for patients with SATP, along with the corresponding survival probability.

Figure 1. (A) Least absolute shrinkage and selection operator (LASSO) regression coefficient path. (B) Cross-Validation (CV) LASSO regression coefficient path. The coefficient path of the LASSO regression illustrates the variation in each variable’s coefficient as the regularisation parameter, λ, increases. Through cross-validation, the CV LASSO regression coefficient path shows how the coefficients behave after λ is changed. Both paths provide an in-depth understanding of how regularisation affects variable selection and coefficient estimation in the LASSO regression model.

Table 2. Multivariable Cox regression analysis of mortality risk in patients with SATP.

Characteristics	B	SE	HR	CI	Z	P value
Age	0.015	0.004	1.015	1.007 to 1.023	3.558	<0.001
Pneumonia	0.245	0.119	1.278	1.012 to 1.613	2.064	0.039
Platelet	−0.002	0.001	0.998	0.996 to 1	−2.552	0.011
RDW	0.046	0.021	1.047	1.005 to 1.09	2.225	0.026
Total bilirubin	0.029	0.007	1.029	1.015 to 1.044	4.094	<0.001
Urea nitrogen	0.004	0.002	1.004	1 to 1.008	1.986	0.047
Lactate	0.133	0.027	1.142	1.083 to 1.204	4.923	<0.001
SPO2	−0.069	0.014	0.934	0.909 to 0.959	−5.022	<0.001
Ventilation	0.789	0.136	2.2	1.687 to 2.871	5.813	<0.001

B, regression coefficient; RDW, red cell distribution width; SATP, sepsis-associated thrombocytopenia; SPO₂, oxygen saturation; Z, Wald Z-statistic.

Figure 2. Nomogram for predicting 3-month mortality in patients in the ICU with SATP model performance. ICU, intensive care unit; RDW, red cell distribution width; SATP, sepsis-associated thrombocytopenia; SPO₂, oxygen saturation.

The predictive performance of the nomogram was evaluated using ROC analysis (figure 3). In the training set, the nomogram achieved an AUC of 0.868 (95% CI: 0.794 to 0.942) for predicting 90-day mortality. In the validation set, the AUC was 0.836 (95% CI: 0.681 to 0.991), indicating good discriminative ability. A calibration curve is presented in figure 4 to demonstrate the predictive model’s performance in both the training set (A) and the validation set (B). The calibration curves visually assess the agreement between predicted probabilities and observed probabilities. The DCA curves for both cohorts are displayed in figure 5, clearly demonstrating net clinical benefit. Compared with individual prognostic factors, the nomogram demonstrated a greater overall net benefit.

Figure 3. ROC AUC values and bootstrap validation. (A) ROC curve for the 3-month training set. (B) ROC curve for the 3-month validation set. AUC, area under the curve; ROC, receiver operating characteristic.

Figure 4. Calibration curve analysis for the training set (A) and the internal test set (B). CSS, Calibration Curve Coordinate System.

Figure 5. Time-dependent decision curve analysis for the training set (A) and validation set (B).

The nomogram depicts a descriptive chart that includes risk scores for each predictive variable. Higher scores are intrinsically associated with increased predicted mortality. Using this schematic, patients were scored and categorised into high-risk and low-risk tiers. In order to evaluate the consistency of these cohorts, the Kaplan-Meier survival curves presented in figure 6 were used. Undoubtedly, in both data partitions, the mortality risk index was observed to be lower in the low-risk group than in the high-risk group.

Figure 6. Rationale analysis: Kaplan-Meier survival curves for high-risk and low-risk groups. (A) The training set. (B) The validation set.

Discussion

SATP is a frequent complication in the ICU, with thrombocytopenia occurring in up to 55% of patients with septic shock.¹¹ This study, conducted using data from the MIMIC-IV, identified nine prognostic factors that were significantly associated with outcomes. Furthermore, a nomogram model was developed to predict 90-day mortality. The model incorporated vital signs, laboratory results, comorbidities and treatment measures obtained on admission, all of which are readily accessible. In both internal and external validation cohorts, it demonstrated exceptional predictive performance, indicating broad applicability and therapeutic utility. The model was not only predictive but also capable of identifying and quantifying key clinical risk factors influencing SATP mortality, providing clinicians with actionable insights. The Cox model framework, coupled with the visual nomogram, offers a transparent and clinically interpretable representation of these relationships, facilitating direct clinical application and understanding of how each factor contributes to the overall risk estimate.

Our study suggested that patients with SATP may have decreased mortality when their platelet counts are higher. Multiple published articles have consistently demonstrated that thrombocytopenia is significantly linked to poor prognosis in patients and correlates closely with its severity.¹² A multicentre observational study in UK ICUs¹³ revealed that critically ill patients with severe thrombocytopenia (platelet count<50×10⁹/L) had a mortality rate of 35.4%, with the severity of thrombocytopenia significantly correlating with patient mortality. Previous research¹⁴ has indicated that in patients with sepsis, mortality rates are notably higher among those with a platelet count<50×10⁹ /L compared with those with a platelet count<100×10⁹/L, which aligns with our findings. Platelets play a pivotal role in the inflammatory response¹⁵ by enhancing the effects of inflammation through various mechanisms, such as the release of inflammatory mediators, promotion of leucocyte adhesion and chemotaxis and synergistic interactions with other cells. These actions aid in the clearance of pathogens and the repair of damaged tissues. Also, platelets can trigger an acute phase response to infection,¹⁶ leading to the production of proteins, such as complement proteins, fibrinogen and C-reactive proteins. These proteins disrupt microbial growth and facilitate coagulation.^{17 18}

Hyperbilirubinaemia is a well-known complication of sepsis, and elevated bilirubin levels may contribute to the disease process. Bilirubin in the bloodstream can induce haemolysis of red blood cells.¹⁹ Elevated bilirubin levels can stimulate oxidative stress, reduce cell survival and promote apoptosis, as demonstrated in cultured cells.²⁰ Moreover, bilirubin can induce inflammatory responses, which are further exacerbated when cells are simultaneously exposed to lipopolysaccharides. Studies have indicated that hyperbilirubinaemia is associated with an overall poor prognosis in critically ill patients.²¹

Previous studies have identified BUN as an independent risk factor for mortality in patients with sepsis. In patients with sepsis, the 30-day mortality rate gradually increases with higher BUN levels.²² BUN is the primary end product of protein metabolism in the human body, mainly excreted through the kidneys, and is also a marker of renal impairment. In patients with sepsis, the rate of protein catabolism is significantly increased²³ and is often accompanied by acute kidney injury,²⁴ both of which can lead to elevated BUN levels in patients with sepsis.²⁵

RDW has been gradually applied as a prognostic indicator for various diseases. High RDW is associated with poor outcomes in critically ill patients, particularly in those with septic shock.²⁶ Previous studies have shown that elevated RDW is linked to increased mortality in patients with sepsis.²⁷ After the onset of sepsis, inflammatory mediators trigger endothelial cell activation, increase vascular permeability, disrupt coagulation and cause microcirculatory dysfunction, all of which are associated with elevated RDW.²⁸ Oxidative stress is a key feature of sepsis, damaging cellular components and increasing cell turnover, including red blood cells, making RDW a potential marker of this condition.²⁹

Serum lactate is considered a biomarker for sepsis prognosis, with elevated serum lactate levels positively correlated with sepsis mortality.³⁰ The recent Sepsis-3 guidelines recommend that persistent serum lactate levels exceeding 2 mmol/L, despite adequate fluid resuscitation, should be used as a new clinical criterion for the definition of septic shock.¹ Studies have shown that high serum lactate concentrations may predict mortality, while a decrease in lactate levels is associated with improved clinical outcomes.^{31 32}

Older patients with sepsis are often associated with a worse prognosis. In a study of over 100 000 patients with sepsis, the average patient age was 71 years.³³ These patients typically present with multiple comorbidities and functional limitations.³⁴ Increased frailty is often associated with higher mortality in older individuals.³⁵ Additionally, multiple comorbidities and polypharmacy in this population may contribute to the development of thrombocytopenia.³⁶

Community-acquired pneumonia has a high incidence and mortality and is a leading cause of sepsis.³⁷ Among these patients, those requiring vasopressor support have a mortality rate as high as 50%.³⁸ Thrombocytopenia is frequently observed in patients with pneumonia and is significantly associated with a worsened prognosis.³⁹ Extensive alveolar damage can lead to the destruction of pulmonary capillaries and the lysis of megakaryocytes, resulting in thrombocytopenia.⁴⁰ Patients with normal platelet counts at admission who develop thrombocytopenia during their ICU stay have lower survival rates compared with those without thrombocytopenia.⁴¹

Mechanical ventilation is a primary life-support measure for critically ill patients. Proper management of sepsis should include ventilation support to mitigate lung injury.² However, injudicious ventilation strategies can lead to lung overdistension, provoke systemic inflammation⁴² and subsequently cause organ dysfunction, contributing to poorer outcomes and reduced survival rates.⁴³

Our study found that higher SPO₂ was associated with better patient outcomes. A study on oxygenation targets in patients with sepsis suggested that maintaining oxygenation at higher-than-conventional levels might improve patient outcomes.¹⁷ Another study found that patients treated with a lower oxygenation target had notably higher mortality compared with those treated with a higher target.¹⁸ This might be attributed to the fact that sepsis can cause hypoxaemia by impairing cellular oxygen transport and utilisation, thereby increasing patient mortality. However, it is important to acknowledge a significant limitation regarding SpO₂ as a predictor. The MIMIC-IV database had variable completeness of fractional inspired oxygen (FiO₂) data recorded within the first 24 hours, preventing us from reliably incorporating PaO₂/FiO₂ into our model construction. Future studies aiming to refine prognostic models in patients with SATP should prioritise the inclusion of more robust measures of oxygenation. Despite this limitation, SpO₂ remained a significant predictor in our model. Regardless of FiO₂ requirements, they may still represent a group with less severe respiratory compromise or better physiological reserve in the context of SATP. Additionally, our study incorporated the patient’s mechanical ventilation status and arterial PaO_₂, which, to some extent, addresses the patient’s oxygenation needs.

In summary, monitoring the aforementioned risk factors and accurately assessing patient prognosis early, as guided by our model, is crucial. This guidance allows us to implement appropriate interventions, thereby improving the prognosis of patients with SATP. Compared with previous studies, this study offers several advantages. First, it focuses specifically on patients in the ICU and is applicable to all individuals with SATP. The broader inclusion criteria for adult patients with SATP enhance the predictive model’s generalisability. Additionally, this study incorporated a wide array of independent risk factors, including vital signs, laboratory results, comorbidities and treatment measures, all readily obtainable soon after admission, ensuring the model’s simplicity and timely application. Furthermore, the AUC and calibration curves demonstrated that the nomogram exhibited favourable discriminative ability and calibration, suggesting substantial potential for clinical application. However, our findings are subject to several limitations. First, since the data was extracted from the MIMIC-IV database, despite the nomogram model being validated with the validation cohort, further validation using multicentre data is required. Second, the model is not applicable to patients with missing variables; for instance, lactate levels are typically measured within 1 hour of ICU admission per sepsis guidelines, whereas bilirubin testing might occur less frequently. Additionally, there may be variables not included in the model that could also influence in-hospital mortality in patients with SATP.

Conclusions

In summary, a quick and affordable nomogram was created in this study to predict the 90-day mortality rate for inpatients with SATP in the ICU. The nomogram incorporated nine independent risk factors: age, platelet count, RDW, total bilirubin, BUN, lactate, SPO₂, mechanical ventilation support and the presence of pneumonia. Validation has underscored the model’s robust performance in discrimination, calibration and clinical utility. However, to enhance and increase the generalisability of our established nomogram, prospective cohort studies as well as validation targeting diverse populations are necessary.

Data availability statement

Data are available in a public, open access repository.