Machine learning for the early prediction of sepsis patients in the intensive care unit (ICU) based on clinical data

Abstract

Objective

This study aimed to develop and validate machine learning models to predict 28-day mortality in sepsis patients admitted to the intensive care unit.

Methods

Initial clinical data from sepsis patients at the time of hospital admission including demographic characteristics, biochemical markers, infection sites, common comorbidities, and scoring systems were used to predict 28-day mortality of sepsis. Least absolute shrinkage and selection operator regression was applied to identify the most relevant predictive variables. After comparing seven algorithms-adaptive boosting, logistic regression, random forest (RF), K-nearest neighbors, Gaussian Naive Bayes, multilayer perceptron, and decision tree-we rebuilt the prediction model using the best-performing one. The model's performance was evaluated using multiple metrics, including the area under the curve (AUC), sensitivity, specificity, accuracy, positive and negative predictive values, F1 score, kappa statistic, and clinical decision curve analysis. Finally, the interpretability of the best-performing model was evaluated using the SHAP package.

Results

Seven critical features were screened including platelet distribution width to count ratio, mean platelet volume, serum creatinine, lactate, D-dimer, APACHE II score, and respiratory system infection. Among the seven algorithms, RF outperformed the others significantly. After training with the best-performing algorithm, the AUCs of the model in the training and validation sets were 1.0 and 0.933, respectively, and the model also performed well in the test set (AUC = 0.900, sensitivity = 0.742, specificity = 0.902, accuracy = 0.841, F1 score = 0.780).

Conclusions

A 28-day mortality in sepsis patients can be accurately predicted at an early stage using a machine learning model based on routinely collected clinical data.

Introduction

Sepsis is an inflammatory condition characterized by a dysregulated host response to infection, leading to life-threatening multiple organ dysfunction. ¹ In the United States, sepsis is the most common cause of in-hospital death and accounts for more than US$24 billion in annual healthcare costs, remaining a formidable public health challenge.^2,3 Although significant advances have been made in scientific technologies, sepsis remains a leading cause of mortality and imposes a substantial economic burden on intensive care unit (ICU) patients, particularly in developing countries such as China. Therefore, early and dynamic prediction of sepsis and timely implementation of appropriate treatment may improve clinical outcomes.

In recent years, artificial intelligence has been increasingly applied to predict mortality risk in a wide range of diseases. Machine learning, a subset of artificial intelligence, is widely used to develop disease prediction models and is effective for early and accurate prediction of mortality in patients with sepsis.^4,5 Despite promising research results, many prediction models still require further validation and practical testing before they can be applied in clinical practice.

The aim of this study was to develop an interpretable model to predict 28-day mortality risk in ICU patients with sepsis, using clinical data from a Chinese cohort to comprehensively and objectively evaluate the associated risk factors. In addition, predictive factors were selected using the least absolute shrinkage and selection operator (LASSO), and the Shapley Additive Explanations (SHAP) method was employed to intuitively interpret the contribution of each risk factor to individual predictions in the machine learning models.

Materials and methods

Patients and collected variables

This study included 819 patients who met the Sepsis-3.0 criteria and were admitted to the ICU of the Affiliated Hospital of Chengde Medical University between 2022 and October 2025. All patients received management in accordance with international guidelines. Each cohort was divided into survival and nonsurvival groups based on 28-day mortality outcomes. The inclusion criteria were as follows: (a) a diagnosis of sepsis based on the Third International Consensus Definitions for Sepsis and Septic Shock (Sepsis-3); (b) age ≥18 years; and (3) patients admitted to the ICU with complete clinical data. The exclusion criteria were as follows: (1) presence of autoimmune diseases; (2) diagnosis of malignant tumors; (3) current treatment with glucocorticoids or immunosuppressants; and (4) pregnancy. A flowchart of the study is shown in Figure 1. The study was approved by the Ethics Committee of the Affiliated Hospital of Chengde Medical University (Approval number: CYFYLL2023401). The need for written informed consent was waived because of the retrospective nature of the study.

Figure 1.

Flowchart of sepsis selection.

Clinical data collection

The clinical data of all the enrolled ICU patients with sepsis were obtained from hospital medical records: (1) sex and age; (2) common comorbidities, including hypertension, cerebrovascular disease, diabetes, coronary artery disease and chronic pulmonary disease; (3) site of infections, including respiratory, gastrointestinal, genitourinary, hepatobiliary, skin and soft tissue, and other site; (4) biochemical indicators on admission, including C-reactive protein, white blood cell count, hemoglobin, platelet (PLT) count, platelet distribution width (PDW), platelet distribution width to count ratio (PCR), mean platelet volume (MPV), neutrophil count, lymphocyte count, monocyte count, fibrinogen level, albumin level, D-dimer, total bilirubin, alanine aminotransferase, aspartate aminotransferase, lactate dehydrogenase, serum creatinine (Scr), blood urea nitrogen, procalcitonin and lactate (Lac) levels; and (5) a scoring system, including the APACHE II score.

Statistical methods

All statistical analyses were conducted using R Studio (3.5). Patients with sepsis were randomly divided into training and validation sets in a 7:3 ratio. Categorical variables are presented as frequencies and percentages, with group differences assessed using the χ² test or Fisher's exact test, as appropriate. Continuous variables are reported as medians with interquartile ranges (IQRs), and comparisons between groups were performed using the Wilcoxon rank-sum test. Crucial features were screened by the LASSO regression method from 35 first recorded variables at admission.

This study employed multiple machine learning algorithms to process the data. The algorithms used include adaptive boosting (AdaBoost), logistic regression (LR), random forest (RF), K-nearest neighbors (KNN), Gaussian Naive Bayes (GNB), multilayer perceptron (MLP), and decision tree (DT). The predictive performance of each model was evaluated using the area under the curve (AUC). The forest plot displays the receiver operating characteristic (ROC) curve results for each algorithm in predicting mortality among the sepsis patients. Decision curve analysis was conducted to evaluate the clinical utility of the predictive models by quantifying their net benefit across a range of threshold probabilities.

Results

Patient characteristics

A total of 819 adult patients diagnosed with sepsis were included in the final cohort based on predefined inclusion and exclusion criteria. The median age of the sepsis patients was 68 years (IQR, 57–76), and 60.2% were male. Patients were randomly assigned to a training set (n = 573) and a validation set (n = 246) in a 7:3 ratio. There were no statistically significant differences in baseline characteristics, comorbidities, or laboratory parameters between the two groups (all P > 0.05). Detailed baseline data for both sets are provided in Table 1.

Table 1.

Baseline characteristics of patients with sepsis between in training and validation sets.

	Training set (n = 573)			Validation set (n = 246)
Variables	Survival group (n = 354)	Death group (n = 219)	P-value	Survival group (n = 155)	Death group (n = 91)	P-value	P-value among sets
Age (years)	67 (57, 76)	69 (59, 78)	0.320	68 (57, 76)	67 (57, 76)	0.843	0.632
Sex, n (%)			0.010			0.834	0.542
Female	158 (44.63)	74 (33.79)		60 (38.71)	34 (37.63)
Male	196 (55.37)	145 (66.21)		95 (61.29)	57 (62.64)
CRP (mg/L)	132.22 (77.36, 211.55)	151.88 (68.89, 218.80)	0.322	136.00 (75.63, 197.63)	130.93 (72.73, 197.26)	0.848	0.476
WBC (×109/L)	11.29 (6.99, 16.18)	11.64 (6.30, 16.45)	0.547	12.00 (7.30, 16.33)	10.99 (5.06, 16.77)	0.553	0.899
Hemoglobin(g/L)	104.00 (91.00, 122.00)	105.00 (84.00, 123.00)	0.535	107.00 (88.00, 125.00)	99.00 (85.00, 116.00)	0.058	0.871
PLT (×109/L)	103.00 (74.00, 156.00)	59.00 (35.00, 110.00)	<0.001	95.00 (71.00, 138.00)	50.00 (30.00, 99.00)	<0.001	0.168
PDW (%)	13.20 (11.40, 16.30)	15.60 (12.80, 18.20)	<0.001	13.80 (11.10, 16.80)	15.10 (12.40, 17.60)	<0.001	0.879
PCR (%)	12.64 (8.34, 21.18)	28.33 (12.76, 45.31)	<0.001	14.61 (8.24, 21.22)	33.81 (12.81, 60.74)	<0.001	0.235
MPV (fL)	10.80 (10.10, 11.70)	13.20 (11.50, 14.00)	<0.001	10.90 (10.00, 12.10)	13.20 (11.40, 14.30)	<0.001	0.908
Neu (%)	89.00 (84.50, 92.40)	89.10 (81.50, 92.90)	0.968	89.64 (82.70, 93.20)	89.60 (81.20, 92.10)	0.302	0.828
Lym (%)	5.54 (3.50, 9.00)	5.80 (3.30, 11.70)	0.373	5.30 (3.00, 10.10)	6.20 (4.02, 11.20)	0.070	0.700
Mono (%)	4.20 (2.70, 6.20)	4.10 (2.10, 6.40)	0.053	3.80 (2.50, 6.10)	3.20 (2.20, 5.50)	0.121	0.321
Alb (g/L)	27.13 ± 5.12	25.21 ± 5.91	<0.001	26.05 ± 5.36	24.14 ± 5.30	0.007	0.060
Tbil (µmol/L)	23.42 (14.00, 35.30)	32.70 (20.12, 55.15)	<0.001	22.30 (13.46, 37.97)	37.81 (20.91, 80.18)	<0.001	0.415
ALT (U/L)	33.00 (21.00, 77.00)	62.16 (29.00, 163.00)	<0.001	32.00 (18.81, 67.80)	74.73 (30.98, 230.95)	<0.001	0.837
AST (U/L)	51.00 (28.27, 85.50)	120.98 (52.23, 338.25)	<0.001	47.38 (29.59, 90.56)	116.50 (52.00, 435.44)	<0.001	0.565
LDH (U/L)	293.00 (224.00, 412.00)	606.71 (336.00, 1230.00)	<0.001	304.00 (219.80, 454.67)	585.91 (340.00, 1300.00)	<0.001	0.576
Scr (µmol/L)	142.96 (113.52, 177.90)	219.00 (151.42, 315.92)	<0.001	143.26 (115.26, 175.20)	215.00 (153.50, 301.00)	<0.001	0.895
BUN (mmol/L)	12.21 (8.96, 15.93)	17.83 (12.21, 27.53)	<0.001	11.89 (9.10, 15.39)	16.49 (10.90, 25.61)	<0.001	0.364
Fib (G/L)	4.73 (2.98, 5.78)	5.42 (2.98, 6.95)	0.012	4.68 (2.82, 5.71)	5.55 (3.29, 6.90)	0.006	0.860
D_dimer (µg/mL)	5.05 (2.37, 8.96)	13.43 (6.49, 22.20)	<0.001	5.51 (2.33, 9.83)	15.51 (6.12, 30.12)	<0.001	0.496
PCT (ng/mL)	17.44 (3.86, 52.02)	20.09 (4.33, 83.05)	0.083	11.85 (2.49, 52.86)	21.80 (3.95, 59.49)	0.267	0.378
Lac (mmol/L)	3.40 (2.60, 4.62)	7.60 (4.10, 12.20)	<0.001	3.58 (2.60, 4.83)	8.30 (3.60, 12.30)	<0.001	0.949
Site of infection
Respiratory, n (%)	91 (25.71)	109 (49.77)	<0.001	36 (32.23)	37 (40.66)	0.004	0.146
Gastrointestinal, n (%)	110 (31.07)	32 (14.61)	<0.001	56 (36.13)	14 (15.39)	<0.001	0.271
Genitourinary, n (%)	44 (12.43)	25 (11.42)	0.717	25 (16.13)	10 (11.00)	0.265	0.389
Hepatobiliary, n (%)	50 (12.14)	42 (11.86)	0.306	26 (16.77)	15 (16.48)	0.953	0.779
Skin, n (%)	25 (7.06)	13 (5.08)	0.599	8 (5.16)	5 (5.50)	0.910	0.465
Other, n (%)	20 (5.65)	18 (8.22)	0.230	4 (2.58)	13 (14.27)	<0.001	0.884
Comorbidity
Hypertension, n (%)	140 (39.55)	85 (38.81)	0.861	52 (33.55)	21 (23.08)	0.083	0.090
CAD, n (%)	63 (17.80)	52 (23.74)	0.084	27 (22.31)	20 (22.00)	0.200	0.190
Diabetes, n (%)	91 (25.71)	52 (23.74)	0.598	35 (22.58)	22 (24.18)	0.775	0.586
CVD, n (%)	81 (22.88)	54 (24.67)	0.626	20 (12.90)	15 (16.48)	0.438	0.300
COPD, n (%)	15 (4.24)	10 (4.57)	0.851	3 (1.94)	4 (4.40)	0.263	0.304
Scoring system
APACHE II score	21.00 (18.00, 28.00)	28.00 (22.00, 35.00)	<0.001	21.00 (18.00, 26.00)	26.00 (22.00, 32.00)	<0.001	0.169

CRP: C-reactive protein; WBC: white blood cell count; PLT: platelet; PDW: platelet distribution width; PCR: platelet distribution width to count ratio; MPV: mean platelet volume; Neu: neutrophil; Lym: lymphocyte; Mono: monocyte; Alb: albumin; Tbil: total bilirubin; ALT: alanine aminotransferase; AST: aspartate aminotransferase; LDH: lactate dehydrogenase; Scr: serum creatinine; BUN: blood urea nitrogen; Fib: fibrinogen; PCT: procalcitonin; Lac: lactate; CAD: coronary artery disease; CVD: cerebrovascular disease; COPD: chronic obstructive pulmonary disease.

Predictive indicators selected from LASSO regression

LASSO regression was applied to the initial training dataset for variable selection. To prevent data leakage and optimistic performance estimation, all data preprocessing steps—including missing value imputation, feature scaling, and feature selection—were performed within each training fold of the cross-validation procedure. Specifically, for each fold, preprocessing parameters were learned exclusively from the training data and subsequently applied to the corresponding validation data. Feature selection was also conducted independently within each training fold. Using 10-fold cross-validation, seven variables associated with the prognosis of sepsis patients were identified, as shown in Figure 2. In this study, predictors were selected based on the λ value corresponding to one standard error above the minimum mean squared error, a commonly used criterion to enhance model simplicity and generalizability. At this threshold, seven nonzero coefficients were retained.

Figure 2.

Texture feature selection using least absolute and selection operator (LASSO).

Comparison of multiple machine learning models

Data classification was performed using seven machine learning algorithms. Considering all performance indicators, RF was identified as the most robust and accurate algorithm, with AUC of 1.0 in the training set and 0.933 in the validation set, as shown in Figure 3(a) and (b). However, the presence of overfitting in the training set suggests that model performance should be interpreted with caution, as generalizability to new data may be affected. In addition, the cut-off value, sensitivity, specificity, accuracy, positive predictive value, negative predictive value, F1 score, and kappa value were 0.550, 0.835, 0.937, 0.894, 0.905, 0.887, 0.869, and 0.780 in the validation set, as shown in Table 2. According to the pairwise DeLong test, RF demonstrated significantly superior discrimination performance compared with all other machine learning models. The calibration curve reflects the accuracy of each model's predictions; the closer the fitted line is to the reference line, the smaller the value in brackets, as shown in Figure 3(c). The clinical decision curve depicts the net benefit of each model, as shown in Figure 3(d). Precision–recall curves of machine learning models for predicting 28-day mortality in patients with sepsis in the validation set, as shown in Figure 3(e).

Figure 3.

Comparison of seven machine learning models in training and validation sets (a, b). Calibration curve of validation models built by machine learning models (c). Decision curve analysis of seven models plotting the net benefit at different threshold probabilities (d). Precision–recall curves of machine learning models in the validation set (e).

Table 2.

Multimodel classification results in the validation set.

Model	AUC (95% CI)	Cut-off	Sensitivity (%)	Specificity (%)	Accuracy	Positive predictive value	Negative predictive value	F1 score	Kappa
AdaBoost	0.890 (0.848–0.933)	0.497	0.816	0.811	0.813	0.757	0.859	0.785	0.620
Logistic	0.883 (0.838–0.928)	0.431	0.748	0.874	0.821	0.811	0.828	0.778	0.629
Random forest	0.933 (0.901–0.965)	0.550	0.835	0.937	0.894	0.905	0.887	0.869	0.780
KNN	0.884 (0.840–0.927)	0.400	0.796	0.839	0.821	0.781	0.851	0.788	0.634
GNB	0.888 (0.845–0.931)	0.286	0.806	0.867	0.841	0.814	0.861	0.810	0.674
MLP	0.880 (0.835–0.925)	0.397	0.728	0.867	0.809	0.798	0.816	0.761	0.603
Decision tree	0.824 (0.775–0.873)	1.000	0.788	0.859	0.830	0.802	0.849	0.795	0.649

AUC: area under the curve; CI: confidence interval; AdaBoost: adaptive boosting; KNN: K-nearest neighbors; GNB: Gaussian Naive Bayes; MLP: Multilayer Perceptron.

Optimal predictive model

A multimodel comparison showed that RF performed best. A test set comprising 10% of cases was randomly selected from the overall sample. The AUC for the test set was 0.900 using the RF model, as shown in Figure 4(a). The calibration curve illustrates the agreement between the RF model's predicted outcomes and the actual observations, thereby reflecting its predictive accuracy, as shown in Figure 4(lb). As shown in the learning curve (Figure 4(c)), the RF model maintained a high AUC in both training and validation sets, suggesting good model stability and generalization ability. The clinical decision curve illustrates the net clinical benefit of the RF model in the test set, as shown in Figure 4(d). The ROC curve of the APACHE II score for predicting mortality in patients with sepsis in the test set. The AUC was 0.707 (95% confidence interval (CI): 0.674–0.739), as shown in Table 3.

Figure 4.

The ROC results of the model were generated using the RF model for the test set (a). Calibration plots of models built by RF model (b). Learning curve of the RF model (c). Calibration curve of validation models built by RF model (d). Decision curve analysis of the RF model (e). The ROC curve of the APACHE II score for predicting mortality in patients with sepsis of test set (f). ROC: receiver operating characteristic; RF: random forest.

Table 3.

Comparison of the discrimination performance of the RF model and the APACHE II score for the test set.

Model	AUC (95% CI)	Cut-off	Sensitivity (%)	Specificity (%)	Accuracy
RF	0.900(0.833–0.967)	0.450	0.842	0.902	0.841
APACHE II score	0.707(0.674–0.739)	23	0.739	0.595	0.646

RF: random forest; AUC: area under the curve; CI: confidence interval.

Interpretation of the model

The SHAP analysis was conducted on the validation dataset of the RF model to interpret the prediction of mortality in patients with sepsis. The variable importance ranking shows that MPV is the most influential predictor, followed by D-dimer, Lac level, Scr, PCR, APACHE II score, and respiratory system infection, as shown in Figure 5(a) and (b).

Figure 5.

The weights of variables importance (a). The Shapley Additive exPlanations (SHAP) values (b). The sepsis patients identified by the random forest model in the validation dataset (c).

In the SHAP diagram for the validation set, the color scale from blue to red represents an increase in the absolute value of the horizontal coordinate. A large negative value on the horizontal axis indicates a higher likelihood of a negative prediction, whereas a large positive value indicates a higher likelihood of a positive prediction. Red bars, such as elevated D-dimer levels, a high APACHE II score, and respiratory system infection, indicate features that increase the predicted probability of mortality. In contrast, blue bars, such as lower MPV levels, represent features that decrease this probability, as shown in Figure 5(c).

Discussion

Sepsis remains a leading cause of morbidity and mortality worldwide and represents an increasing public health threat in China, underscoring the urgent need for early, personalized intervention and treatment strategies. Many existing studies have relied on large public datasets, such as MIMIC-III and eICU. Models based on single-center data may be subject to scrutiny to some extent. ⁶ To our knowledge, this is the first study to develop a machine learning-based clinical prediction model for ICU sepsis mortality using limited, readily available clinical data. This study employed machine learning techniques to develop a predictive model incorporating seven predictors of sepsis events identified by the RF algorithm. The model achieved an AUC of 0.933 in the validation set, indicating excellent discriminative performance. The RF model outperformed other models, including AdaBoost, LR, KNN, GNB, MLP, and DT, and conventional severity score. The RF model was prioritized for its efficiency in handling high-dimensional data and capturing nonlinear interactions, as demonstrated in previous critical care studies.^7,8 It has been extensively used to predict in-hospital mortality in sepsis patients and may assist clinicians in decision-making. Wang et al. developed a predictive model using the RF method, incorporating 20 predictors and data from 4449 sepsis patients. ⁹ For the external validation, the AUC of the model was 0.91.

To improve the interpretability of the predictive model, we applied the SHAP method to quantify the contribution of each variable to the model's mortality predictions in sepsis patients. Several features, including Scr, Lac level, and the APACHE II score, have been identified in previous models predicting mortality in patients with sepsis.^10–12 Importantly, our study identified several previously unreported predictors—PCR, MPV, D-dimer, and respiratory system infection—that were significant characteristics overlooked by traditional predictive models.

Studies have demonstrated that PLT-related parameters, including PDW, PLT, and MPV, serve as important indicators of PLT function and disease severity in patients with sepsis.^13,14 PDW reflects PLT morphological changes and may be associated with PLT function and production rate. ¹⁵ The PCR, which integrates two readily available PLT indices—PDW and PLT count—has been identified as a prognostic biomarker in sepsis. Previous studies have reported its significant value in assessing disease severity in children with sepsis and in predicting clinical outcomes. ¹⁶ An elevated MPV reflects PLT hyperactivity and coagulation activation, and is closely associated with disease severity and poor prognosis in patients with septic shock. ¹⁷ Over the past decade, the role of coagulation in the pathogenesis of inflammation has been increasingly recognized. In addition, sepsis is frequently associated with disseminated intravascular coagulation and microthrombosis. This coagulopathy is characterized by extensive PLT activation, leading to microvascular thrombus formation and elevated D-dimer levels.^18,19

Sepsis exhibits diverse clinical manifestations, with prognosis largely determined by the primary site of infection. Respiratory infections are the most common cause of sepsis and septic shock, and they continue to draw significant attention from both clinicians and researchers.^20,21 In a large retrospective multicenter study, pneumonia accounted for 20.77% of all sepsis-related deaths from 1999 to 2001, increasing to 27.63% from 2020 to 2022. ²² A retrospective cohort study using a national database demonstrated that the site of infection is associated with in-hospital mortality in sepsis patients, with lower respiratory tract infections emerging as the leading cause of death. ²³ Therefore, the anatomical site of infection likely plays a significant role in sepsis-related mortality. Studying current trends in infection sites and their outcomes, and developing targeted preventive measures for the most common or high-risk sites, are essential for optimizing intensive care resource allocation and informing public health strategies.

This study has several limitations. First, it was based on a retrospective cohort with a relatively small sample size, which may limit the generalizability of the findings. Although rigorous cross-validation procedures were applied and all preprocessing and feature selection steps were strictly conducted within each training fold to minimize the risk of data leakage, the relatively high training performance (including a training AUC of 1.0) may still reflect a degree of overfitting inherent to complex machine learning models, particularly in retrospective observational data. Future large-scale, multicenter studies are needed to validate these machine learning models. Second, due to substantial differences in healthcare resources, ICU settings, and laboratory-testing capabilities across regions in China, the findings of this study may be more applicable to patients with sepsis admitted to ICUs within the country. Finally, the absence of an external validation cohort limits the assessment of the clinical applicability of the developed LR model, which warrants further verification.

Conclusion

This study developed a machine learning model that predicts 28-day mortality in sepsis patients using only clinically available data, enabling clinicians to implement preventive measures and improve outcomes in high-risk cases.