Construction and Validation of a Clinical Prediction Model for Sepsis Based on Peripheral Perfusion Index: In-Hospital and 28-Day Mortality Risk Prediction

doi:10.21203/rs.3.rs-4893729/v1

Download PDF

Article

Construction and Validation of a Clinical Prediction Model for Sepsis Based on Peripheral Perfusion Index: In-Hospital and 28-Day Mortality Risk Prediction

https://doi.org/10.21203/rs.3.rs-4893729/v1

This work is licensed under a CC BY 4.0 License

Journal Publication

published 05 Nov, 2024

Read the published version in Scientific Reports →

You are reading this latest preprint version

Background

Sepsis is a clinical syndrome caused by infection, leading to organ dysfunction due to a dysregulated host response. In recent years, its high mortality rate has made it a significant cause of death and disability worldwide. The pathophysiological process of sepsis is related to the body's dysregulated response to infection, with microcirculatory changes serving as early warning signals that guide clinical treatment. The Peripheral Perfusion Index (PI), as an indicator of peripheral microcirculation, can effectively evaluate patient prognosis. This study aims to develop two new prediction models using PI and other common clinical indicators to assess the mortality risk of sepsis patients during hospitalization and within 28 days post-ICU admission.

Methods

This retrospective study analyzed data from sepsis patients treated in the Intensive Care Unit of Peking Union Medical College Hospital between December 2019 and June 2023, ultimately including 645 patients. LASSO regression and logistic regression analyses were used to select predictive factors from 35 clinical indicators, and two clinical prediction models were constructed to predict in-hospital mortality and 28-day mortality. The models' performance was then evaluated using ROC curve, calibration curve, and decision curve analyses.

Results

The two prediction models performed excellently in distinguishing patient mortality risk. The AUC for the in-hospital mortality prediction model was 0.82 in the training set and 0.73 in the validation set; for the 28-day mortality prediction model, the AUC was 0.79 in the training set and 0.73 in the validation set. The calibration curves closely aligned with the ideal line, indicating consistency between predicted and actual outcomes. Decision curve analysis also demonstrated high net benefits for the clinical utility of both models.

Conclusion

The study shows that these two prediction models not only perform excellently statistically but also hold high practical value in clinical applications. The models can help physicians accurately assess the mortality risk of sepsis patients, providing a scientific basis for personalized treatment.

Sepsis

Peripheral Perfusion Index (PI)

In-hospital Mortality

28-day Mortality. Clinical Prediction Model

LASSO Regression

Sepsis is a life-threatening condition caused by a dysregulated host response to infection, resulting in severe organ dysfunction¹. It is a clinical syndrome that encompasses a range of systemic physiological, pathological, and biochemical abnormalities². The mortality rate of sepsis is high, with the severity of the disease progressing rapidly from sepsis to septic shock. Mortality rates for sepsis vary among different populations but are generally estimated to be between 30% and 50%. When patients progress to septic shock, the mortality rate increases even further^3,4. In recent years, the establishment of new sepsis criteria has allowed for more standardized and early identification of sepsis. However, despite these advancements, the mortality rate remains high, making sepsis one of the leading causes of death worldwide^5,6. The research into sepsis has deepened our understanding of its pathophysiology, revealing that it is not merely a battle between the body and pathogens or underlying diseases. Rather, many of the pathological changes are linked to the body's dysregulated response to infection, highlighting the interplay of intrinsic and extrinsic factors in the progression of sepsis⁷.

Among the various factors, the mismatch between macro-circulation and microcirculation hemodynamics in sepsis is a critical area of focus^8–10. In the early stages of sepsis, changes in macro-circulatory hemodynamics are often difficult to detect^11,12. As the disease progresses and pathogens remain unresolved, the host response intensifies, making early detection of sepsis a persistent challenge¹³. In this context, we have observed that changes in microcirculation in peripheral tissues occur earlier than those in macro-circulation and are the last to improve during treatment^14,15. Despite a stable macro-circulation, microcirculatory dysfunction occurs early, with the body compensating to maintain macro-circulation to ensure oxygen supply to vital organs, often at the expense of less critical skin microcirculation^16,17. During the correction of sepsis, maintaining circulatory stability alone does not completely halt disease progression. Only when microcirculatory dysfunction is resolved does the clinical condition improve^18–20. Clinically, we aim to gradually alleviate microcirculatory dysfunction, emphasizing the need to monitor the body’s response and microcirculatory changes during sepsis progression. This also suggests that microcirculation, as a sensitive component of hemodynamics, holds significant potential for assessing and predicting disease outcomes.

There are currently numerous methods for monitoring microcirculation, among which the peripheral perfusion index (PI) is a commonly used clinical indicator. PI is measured using a fingertip sensor that detects changes in blood flow within subcutaneous microvessels, providing a PI value²¹. This method is non-invasive, convenient, and can be performed simultaneously with peripheral blood oxygen saturation measurement. Previous studies have shown that PI can serve as an indicator of microcirculation changes, guiding clinical treatment and offering valuable information about peripheral tissue microcirculation²². Our prior research also found that PI is significantly associated with in-hospital mortality in sepsis patients and can be used to predict organ dysfunction and disease severity²³. In clinical management, precision treatment and personalized management have become new trends. The treatment of sepsis involves not only anti-infection therapy but also the support and monitoring of organ function. Recent studies have shown that early identification and timely treatment are crucial for improving the prognosis of sepsis patients. Hence, new clinical guidelines emphasize early intervention, including prompt administration of antibiotics, fluid resuscitation, and monitoring of organ function¹³. Despite the increasing attention to microcirculation in sepsis patients, reliable tools for predicting sepsis and guiding clinical diagnosis and treatment remain limited in practice. Although our previous research prospectively validated the clinical utility of PI, we believe its predictive efficiency can still be enhanced through the methodology of clinical prediction models.

In this study, we plan to utilize the peripheral perfusion index and other common clinical indicators to develop two new clinical prediction models using clinical data from Peking Union Medical College Hospital. These models will convert multiple indicators into risk probabilities, aiding physicians in more accurately assessing patient conditions and providing a scientific basis for personalized treatment. The introduction of these new clinical prediction models will enhance the precision of clinical management for sepsis patients and offer a broader range of options for personalized therapy. We aim to improve the clinical predictive efficiency of PI through this series of studies, providing more practical and reliable tools for clinical treatment, ultimately increasing the survival rates and quality of life for sepsis patients.

Study Design and Setting

This study employs a retrospective analysis design, encompassing data from sepsis patients treated in the Intensive Care Units (ICU) of the First and Second Departments of Critical Care Medicine at Peking Union Medical College Hospital, China, between December 2019 and June 2023. By thoroughly analyzing patients' clinical data, particularly various clinical predictive variables including the PI, we aim to comprehensively assess the predictive value of these variables for mortality in sepsis patients.

Our research methodology strictly adheres to the principles outlined in the Declaration of Helsinki, drafted in 1975 and revised in 2008, ensuring the protection of research ethics and patient rights. Approval for this research was obtained from the Ethics Committee of Peking Union Medical College Hospital (S-K360), ensuring compliance with all applicable guidelines and regulations. Through a retrospective analysis of extensive data, the study seeks to identify risk factors significantly associated with mortality in sepsis patients, providing crucial tools for clinical prognosis evaluation. The primary objective of this study is to develop two practical clinical prediction models to assess the risk of mortality in sepsis patients during hospitalization and within 28 days after ICU admission. These models will be constructed based on actual clinical data, employing statistical methods such as LASSO regression analysis and Logistic regression analysis. The models will not only visually demonstrate the impact of various risk factors on patient mortality but also provide decision support to physicians, aiding in more precise risk assessment and treatment planning.

Study Subjects

This study included a total of 8,919 critically ill patients treated in the ICU at Peking Union Medical College Hospital between December 2019 and June 2023. Our inclusion criteria were based on the Sepsis-3 diagnostic criteria established in 2016 by the Society of Critical Care Medicine and the European Society of Intensive Care Medicine¹. Patients were required to meet at least one of the following conditions: (1) presence of a confirmed infection source; (2) a Sequential Organ Failure Assessment (SOFA) score of ≥ 2 points. To optimize the representativeness of the sample and the accuracy of the study results, we established several exclusion criteria during the patient screening process. The exclusion criteria were: (1) minors under the age of 18; (2) patients with complex medical histories such as hematologic diseases, malignancies, cirrhosis, or uremia; (3) pregnant or postpartum women; (4) patients with incomplete medical records, i.e., those lacking sufficient documentation to support clinical data analysis; (5) patients lost to follow-up, i.e., those for whom adequate subsequent information could not be obtained during the study period; (6) patients readmitted to the ICU after discharge.

These exclusion criteria helped reduce research bias and allowed us to focus on a patient population with complete data and relatively homogeneous clinical conditions. Ultimately, through a rigorous screening process, we successfully selected 645 sepsis patients who met all inclusion and exclusion criteria.

Data Collection

All study data were collected through the hospital's medical record system. Patient basic information and demographic data included age, gender, diagnosis, height, weight, etc. Hemodynamic indicators comprised the Peripheral Perfusion Index (PI), Systolic Blood Pressure (SBP), Diastolic Blood Pressure (DBP), Mean Arterial Pressure (MAP), Heart Rate (HR), Central Venous Pressure (CVP), and Urine Volume (UV). Serological indicators included White Blood Cell count (WBC), Neutrophil Ratio (NEUT), Lymphocyte Proportion (LYP), Platelet count (PLT), Prothrombin Time (PT), Activated Partial Thromboplastin Time (APTT), Fibrinogen (Fib), D-Dimer (DD), Alanine Transaminase (ALT), Aspartate Transaminase (AST), Direct Bilirubin (Dbil), Total Bilirubin (Tbil), Serum Creatinine (SCR), Blood Urea Nitrogen (BUN), Albumin (Alb), High-sensitivity C-reactive Protein (hsCRP), Cardiac Troponin I (cTnI), Creatine Kinase (CK), N-terminal pro b-type Natriuretic Peptide (NT-proBNP), Procalcitonin (PCT), and Erythrocyte Sedimentation Rate (ESR). Blood gas analysis indicators included pH, Partial Pressure of Oxygen (PO2), Partial Pressure of Carbon Dioxide (PCO2), Central Venous Oxygen Saturation (ScvO2), Tissue Oxygenation Index (TOI), Arterial-venous CO2 Gap (△PCO2), and Lactate (Lac). Clinical assessment scores included the Glasgow Coma Scale (GCS), Sequential Organ Failure Assessment (SOFA), Richmond Agitation-Sedation Scale (RASS), and Acute Physiology and Chronic Health Evaluation II (APACHE II).

All hemodynamic data were obtained from the Philips IntelliVue MP70 Patient Monitor (Philips, USA). Serum samples were tested in the Beijing Union Medical College Hospital laboratory using an automated analyzer, and blood gas analysis samples were processed using the Radiometer Medical AQT90 FLEX (Radiometer, Denmark). Clinical assessment scores were evaluated by at least two specialized physicians, with the average score calculated. All data underwent stringent quality control to ensure reliability.

Statistical Analysis

All statistical analyses were performed using R software (version 4.33; https://www.R-project.org). Missing continuous data, constituting less than 5% of the total, were imputed using the median imputation method^24–26. The study included 645 sepsis patients, who were randomly divided into a training set (387 patients) and a validation set (258 patients) in a 6:4 ratio. Continuous variables were expressed as median (IQR) and compared using the Wilcoxon test. Categorical variables were presented as proportions (%) and compared using the chi-square test and/or Fisher's exact test. To evaluate the predictive ability and accuracy of the clinical prediction models for sepsis patient mortality, ROC curves were plotted, and the AUC, 95% CI, optimal cutoff values, sensitivity, specificity, positive likelihood ratio (PLR), and negative likelihood ratio (NLR) were calculated.

LASSO regression analysis is a method for shrinkage and variable selection in linear regression models^27,28. To obtain a subset of predictors, LASSO regression analysis applies constraints to the model parameters, shrinking some variable regression coefficients to zero, thus minimizing the prediction error of the quantitative response variable. Variables with regression coefficients equal to zero are excluded from the model, while those with non-zero coefficients have the strongest correlation with the response variable. Based on the − 2 log-likelihood and binomial family type metrics, LASSO regression analysis performed in R software centralizes and normalizes the included variables, performing 20-fold cross-validation to select the optimal lambda value. "Lambda.lse" provides a well-performing model with the fewest independent variables²⁹. Additionally, in the training set, LASSO regression analysis is used on clinical baseline indicators to select the best predictors. The selected variables are then incorporated into a multivariate Logistic regression analysis to further evaluate their efficacy by calculating OR values, 95% CIs, and P-values, thereby establishing a nomogram prediction model.

Based on the training and validation sets, several validation methods were employed to evaluate the accuracy of the risk prediction models. The AUC value was used to assess the discriminative ability of the risk nomogram, distinguishing true positives from false positives. Internal validation was performed using the Bootstrap method (1000 resamples), and calibration curves were plotted to evaluate the consistency of the models, specifically assessing the calibration of the in-hospital mortality risk and the 28-day mortality risk after ICU admission for sepsis patients. Decision curve analysis (DCA) was conducted to evaluate the clinical utility of the models^30–32. A two-sided P-value of less than 0.05 was considered statistically significant for all tests.

Participant Characteristics

A total of 645 sepsis patients were included in this study, comprising 412 males and 233 females. These patients were carefully selected to ensure data completeness and the reliability of the study results. Clinical data collected at admission indicated that 92 patients died during their hospital stay. Additionally, 272 patients died within 28 days after ICU admission due to disease progression, voluntary discharge, or transfer to other medical facilities for further treatment. After completing random sampling in a 6:4 ratio, the patients were randomly assigned into groups: 387 were allocated to the training set for in-hospital mortality, and 258 were included in the validation set to verify the accuracy and stability of the models.

To ensure the scientific rigor and systematic approach of the study, all patients' basic data and clinical characteristics were meticulously recorded and summarized in Table 1. These data include, but are not limited to, the patients' age, gender, pre-admission health status, disease severity, and medical history. Based on these data, we developed two clinical prediction models aimed at accurately predicting the in-hospital mortality and 28-day mortality post-ICU admission for sepsis patients. The construction of these models was based on the analysis of various clinical indicators, such as physiological and biochemical parameters, to identify the factors that significantly impact the mortality risk of sepsis patients.

During the model construction process, we paid special attention to the significant differences between the mortality and non-mortality groups. These statistical differences are detailed in Appendix Tables 1 and 2. By comprehensively analyzing these data, we were able to identify key factors affecting patient prognosis. Additionally, we performed correlation significance tests among all relevant feature variables, as shown in Fig. 1, to deepen our understanding of the interactions between different variables. This analysis helped us clarify the relationships between variables, such as positive or negative correlations, and their statistical significance. The significant relationships between some feature variables in the figure indicated the potential for using LASSO regression analysis to address multicollinearity among the feature variables, further selecting independent risk factors as predictors.

Table 1

This table outlines the characteristics of the training and validation sets, highlighting any significant differences between the two groups.
Variables	Total (n = 645)	Training set (n = 387)	Validation set (n = 258)	P value
Age (years)	65.00 (52.00–74.00)	64.00 (51.00–73.00)	66.00 (53.00–75.00)	0.081
Gender, N (%)				0.960
Female	233 (36.00)	139 (35.92)	94 (36.43)
Male	412 (64.00)	248 (64.08)	164 (63.57)
Time (days)	7.00 (3.00–14.00)	7.00 (4.00–15.00)	6.00 (3.00–13.00)	0.075
PI	1.52 (0.86–2.34)	1.51 (0.86–2.35)	1.54 (0.86–2.33)	0.676
HR (bpm)	102.00 (88.00-119.00)	103.00 (87.00-119.00)	101.00 (89.00-118.75)	0.764
SBP (mmHg)	130.00 (111.00-150.00)	130.00 (113.50–150.00)	130.00 (108.25–148.00)	0.261
DBP (mmHg)	67.00 (57.00–76.00)	67.00 (58.00–76.00)	66.00 (56.25–75.75)	0.758
MAP (mmHg)	87.00 (76.00–99.00)	87.00 (76.00-100.00)	87.00 (75.00–99.00)	0.571
CVP (mmHg)	8.00 (7.00–10.00)	8.00 (7.00–10.00)	8.00 (6.00–10.00)	0.939
pH	7.40 (7.34–7.45)	7.40 (7.33–7.45)	7.40 (7.34–7.46)	0.372
PO₂ (mmHg)	98.80 (77.90–127.00)	94.50 (76.55–125.00)	101.00 (79.55-130.75)	0.174
PCO₂ (mmHg)	37.80 (33.00-42.40)	37.80 (33.30–43.00)	37.95 (32.63–42.08)	0.358
ScvO₂ (%)	75.35 (70.20–80.40)	75.35 (70.40–80.00)	75.35 (69.8-80.65)	0.188
TOI (mmHg)	288.00 (200.00-396.70)	285.70 (194.75-390.95)	293.70 (209.48-409.35)	0.288
△PCO₂ (mmHg)	4.25 (2.90–5.70)	4.25 (2.80–5.50)	4.25 (3.00–6.00)	0.408
Lac (mg/dL)	1.80 (1.20–3.20)	1.90 (1.20–3.20)	1.70 (1.10–3.10)	0.159
PCT (ng/mL)	4.30 (0.82-19.00)	4.20 (0.83-19.00)	4.55 (4.55-19.00)	0.690
WBC (*10^9/L)	11.25 (6.98–16.31)	11.61 (7.34–16.31)	10.93 (6.53–16.28)	0.284
NEUT (*10^9/L)	9.97 (5.71–14.66)	10.16 (6.07–14.50)	9.32 (5.48–14.76)	0.308
LYP (%)	6.80 (3.90–11.20)	6.50 (3.80-10.95)	7.15 (4.03–11.20)	0.432
PLT (*10^9/L)	146.00 (92.00-223.00)	150.00 (91.00-224.00)	143.5 (92.25–221.50)	0.443
PT (s)	14.80 (13.50–16.70)	14.90 (13.50–17.10)	14.60 (14.60–16.20)	0.23
APTT (s)	32.80 (28.60–38.70)	32.80 (28.50-39.95)	33.20 (28.70-38.08)	0.882
FIB (g/L)	3.52 (2.54–4.65)	3.52 (2.50–4.69)	3.52 (2.56–4.52)	0.701
D-dimer (mg/L)	4.40 (2.26–8.87)	4.33 (2.18–8.83)	4.52 (2.51–8.90)	0.407
Tbil (µmol/L)	16.60 (10.60–30.20)	16.7 (10.75–30.60)	16.45 (10.53–26.93)	0.588
Dbil (µmol/L)	8.90 (4.70–17.90)	8.90 (4.65–18.05)	8.90 (4.73–15.70)	0.971
ALB (g/L)	30.00 (26.00–33.00)	30.00 (27.00–34.00)	30.00 (26.00–33.00)	0.212
ALT (U/L)	20.00 (13.00–41.00)	20.00 (13.00–35.00)	20.00 (12.25–52.50)	0.178
Cr (µmol/L)	87.00 (61.00-153.00)	90.00 (64.00-158.50)	82.00 (58.25-145.25)	0.084
UREA (mmol/L)	9.31 (5.90–15.00)	9.79 (6.12–16.37)	8.67 (5.41–13.89)	0.085
UV (mL)	2555.00 (1490.00-3970.00)	2540.00 (1402.50–4015.00)	2622.50 (1610.00-3887.50)	0.525
GCS	6 (4–9)	6 (4–9)	6 (4–10)	0.697
SOFA	11 (8–13)	11 (8–13)	11 (8–13)	0.437
APACHE II	20 (15–25)	20 (15–25)	20 (15–25)	0.827

Selection of Independent Risk Factors

To identify factors that significantly predict mortality in sepsis patients, we analyzed 35 relevant feature variables using the LASSO regression method. LASSO regression is a variable selection technique particularly suitable for handling situations with a large number of features and potential multicollinearity³³. Through this method, we successfully identified 9 key factors with strong predictive power for in-hospital mortality. The effectiveness of these factors is demonstrated in Fig. 2(a) and (b). These factors include age, Peripheral Perfusion Index (PI), Systolic Blood Pressure (SBP), Mean Arterial Pressure (MAP), Direct Bilirubin (Dbil), Serum Creatinine (SCR), Urea, Sequential Organ Failure Assessment (SOFA) score, and Acute Physiology and Chronic Health Evaluation (APACHE). Similarly, for predicting 28-day mortality post-ICU admission, we applied LASSO regression analysis to another dataset. From the 35 variables, we identified 5 key factors with strong predictive power. These factors include age, PI, MAP, Neutrophil Count (NEUT), Urea, and APACHE, as shown in Fig. 2(c) and (d).

Development of Prediction Models

Subsequently, we further validated the potential predictors identified by LASSO regression through logistic regression analysis, assessing their correlation with in-hospital mortality and 28-day mortality in sepsis patients. This analysis method allows us to determine the specific strength and direction of each variable's impact on the outcome variable, helping us to better understand how these factors are associated with patient mortality risk.

In the logistic regression model, factors such as age, Peripheral Perfusion Index (PI), Mean Arterial Pressure (MAP), Direct Bilirubin (Dbil), Serum Creatinine (SCR), and Sequential Organ Failure Assessment (SOFA) were found to be closely related to in-hospital mortality, as shown in Table 2. Similarly, logistic regression analysis for 28-day mortality confirmed the importance of age, PI, MAP, Neutrophil Count (NEUT), Urea, and Acute Physiology and Chronic Health Evaluation (APACHE), as shown in Table 3. Using this data, we developed two new clinical prediction models for in-hospital mortality and 28-day mortality post-ICU admission. These models were developed based on the results of the logistic regression analysis, providing a visual and quantitative representation of the impact of these risk factors on patient prognosis. Figures 3(a) and 3(b) illustrate the structure and functionality of these two models, offering practical visualization tools for clinical use.

Table 2

Logistic Regression Analysis of Predictors for In-Hospital Mortality.
Variables	OR (95% CI)	P value
Age (years)	2.97 (1.77–4.98)	< 0.001
PI	0.5 (0.28–0.87)	0.016
DBP (mmHg)	1.2 (0.58–2.47)	0.622
MAP (mmHg)	0.46 (0.23–0.95)	0.035
Dbil (µmol/L)	1.1 (1.00-1.21)	0.044
Cr (µmol/L)	1.22 (1.03–1.46)	0.025
SOFA	1.79 (1.07–2.99)	0.027
APACHE II	1.02 (0.68–0.68)	0.924
UREA (mmol/L)	1 (0.72–1.39)	0.993

Table 3

Logistic Regression Analysis of Predictors for 28-Day Mortality Post-ICU Admission.
Variables	OR (95% CI)2	P value
PI	0.59 (0.40–0.85)	0.005
MAP (mmHg)	0.70 (0.51–0.94)	0.020
WBC (*10^9/L)	0.69 (0.55–0.88)	0.002
APACHE II	1.95 (1.40–2.70)	<0.001
UREA (mmol/L)	1.79 (1.33–2.41)	<0.001

Validation of Prediction Models

During the validation process of the prediction models, various statistical methods were employed to assess their performance and consistency. Firstly, Receiver Operating Characteristic (ROC) curve analysis was used to evaluate the model's ability to distinguish between the occurrence and non-occurrence of events. As shown in Supplementary Fig. 1(a) and 1(b), the area under the ROC curve (AUC) for the in-hospital mortality prediction model was 0.82 in the training set, indicating high predictive accuracy. In the validation set, as illustrated in Supplementary Fig. 1 (c) and 1(d), the model achieved an AUC of 0.73, which, although slightly lower, still demonstrates good predictive performance. Similarly, the 28-day mortality prediction model had an AUC of 0.79 in the training set and 0.73 in the validation set, reflecting stable predictive efficacy.

Additionally, calibration curve analysis further confirmed the models' consistency. The calibration curve illustrates the relationship between the predicted probabilities and the actual occurrence probabilities, with the ideal calibration curve being close to a 45-degree angle. In this study, as shown in Supplementary Fig. 2, the calibration curves for both models closely align with this ideal line, indicating that the predicted values match the actual outcomes very well. This demonstrates the models' reliability and practical applicability.

Decision curve analysis is an essential tool for evaluating the clinical utility of prediction models. This analysis shows the net benefit of using the models to predict outcomes across a range of risk threshold probabilities. In this study, for the in-hospital mortality risk prediction model, the accuracy of assessing the risk of sepsis-related in-hospital mortality was higher when the risk threshold probability ranged from 6–62%, as shown in the training set. In the validation set, this range was 5–51%, as illustrated in Supplementary Fig. 3(a) and 3(b). For the 28-day mortality risk prediction model post-ICU admission, the prediction accuracy was higher when the risk threshold probability ranged from 23–82% in the training set, and from 21–75% in the validation set, as depicted in Supplementary Fig. 3(c) and 3(d). This indicates that within these probability ranges, the prediction models can provide valuable information for clinical decision-making without the concern of overtreatment or missed diagnoses.

Combining these validation results, we can conclude that both prediction models not only perform excellently statistically but also hold high practical value in clinical applications. These models can aid healthcare professionals in more accurately assessing the mortality risk of sepsis patients, enabling more informed treatment decisions. This not only enhances treatment effectiveness but also optimizes resource allocation, significantly improving clinical outcomes for sepsis patients.

Nomogram prediction models offer distinct advantages in statistical analysis and medical research, particularly in handling categorical outcomes and exploring the relationships between predictor variables and outcomes³⁴. These models are well-suited for dealing with categorical outcome variables, such as the mortality or survival status in this study. Nomogram models can effectively estimate the impact of one or more predictor variables on binary or multicategory outcomes. By using logistic functions, they can provide the probability of an outcome occurring in relation to various predictor factors, which is crucial for clinical decision-making³⁵.

In this study, we used nomogram models to analyze the relationships between variables such as age, Peripheral Perfusion Index (PI), Mean Arterial Pressure (MAP), and the mortality risk in sepsis patients. The models can estimate the increase in mortality risk associated with aging or decreasing PI values. This information is vital for physicians when assessing patient risk and developing treatment plans. The outputs of the models are typically expressed in probabilities, making it easy for doctors and researchers to interpret and use the results. Based on these models, patients can be stratified into different risk levels to aid in resource allocation and treatment prioritization. High-risk patients can receive more intensive treatment and monitoring, while low-risk patients can avoid unnecessary medical interventions. Additionally, nomogram prediction models provide a quantitative assessment of patient prognosis, assisting doctors in making more informed clinical decisions. These models can also be used to evaluate the effectiveness of different treatment strategies, thereby guiding the development of personalized therapies.

This study primarily focused on utilizing the Peripheral Perfusion Index (PI) to construct prediction models. The results indicated a significant correlation between PI values and both in-hospital mortality and 28-day mortality post-ICU admission, highlighting the substantial predictive value of this indicator for different mortality outcomes. PI is measured using a photoplethysmographic sensor on a fingertip pulse oximeter, monitoring blood flow in the peripheral microcirculation and calculating the perfusion index of peripheral tissues. It serves as a direct indicator of microcirculatory efficacy^36,37. Sepsis patients typically exhibit microcirculatory dysfunction, which directly impacts tissue oxygenation and cellular metabolism^19,38. When peripheral microcirculatory function is impaired, PI values usually decrease, correlating with disease severity and poor prognosis³⁹. This non-invasive, low-risk, and cost-effective measurement tool makes PI highly practical. Particularly in clinical settings, continuous monitoring with PI provides real-time data for physicians to monitor patients' physiological status.

PI values reflect the state of microcirculatory function throughout the disease course, increasing with improved blood perfusion. In healthy individuals, higher PI values indicate good microvascular function and blood perfusion. However, in sepsis patients, inflammatory responses lead to altered microcirculatory function, causing vasoconstriction, increased permeability, microcirculatory failure, or microthrombosis, resulting in significantly lower PI values. These changes often precede macro-circulatory instability, leading to a mismatch between macro-circulatory and microcirculatory function⁴⁰. Real-time feedback from PI measurements helps physicians determine whether a patient's microvascular perfusion is improving or deteriorating, allowing for adjustments in medication or fluid administration to optimize hemodynamic status. PI is not only useful for real-time monitoring but also serves as a tool for assessing treatment effectiveness and prognosis. In sepsis management, improved microvascular perfusion usually indicates clinical improvement, whereas persistently low PI values suggest poor response to treatment, necessitating further adjustments in therapeutic strategies.

In our previous study, we found that PI values were significantly associated with in-hospital mortality among ICU patients, but not with 28-day mortality post-ICU admission²³. However, in this current study, we found that PI values were significantly correlated with both in-hospital mortality and 28-day mortality post-ICU admission in sepsis patients. We attribute this discrepancy to the sample size. The previous study included 208 patients, of which 117 were sepsis patients, whereas this study included 645 sepsis patients, nearly six times the sample size of the previous study. Sample size plays a crucial role in the reliability of study results. A smaller sample size may lack sufficient statistical power to detect all actual effects, whereas increasing the sample size enhances the ability to detect true associations and reduces the risk of Type I and Type II errors. Moreover, the current study spanned approximately 2.5 years, during which advances in medical technology and changes in clinical settings may have led to variations in treatment approaches. The data from the previous study were collected over a more recent period than the retrospective data of the current study. Differences in ICU management levels over different periods, as well as variations in the types of patients admitted to the hospital, could also affect the detection of the correlation between PI and mortality. For example, during the COVID-19 pandemic, ICUs admitted some COVID-19 patients, resulting in differences in age, gender, underlying conditions, and disease severity among the patient population. These factors may have influenced the observed association between PI and mortality rates.

This study aims to develop two clinical prediction models to assess mortality risk in sepsis patients at different clinical time points: during hospitalization and within 28 days post-ICU admission. This distinction is crucial as it enables physicians to perform precise risk assessments based on the patient's stage of care. The in-hospital mortality model focuses on the short-term mortality risk from ICU admission to discharge (or death), reflecting the patient's acute physiological and pathological state. The 28-day mortality model post-ICU admission provides mid-term prognostic information, helping to understand the patient's survival probability over the longer-term treatment and recovery process. Through these two models, physicians can better understand the severity of sepsis in patients and potential treatment outcomes. When managing sepsis patients, these models can identify high-risk patients who require special attention and urgent intervention. For instance, if a patient shows high risk in the in-hospital mortality model, physicians might adopt more aggressive treatment strategies, such as enhanced monitoring and the use of advanced life support technologies. If the 28-day mortality model predicts a poor outcome, the medical team might discuss long-term prognosis and end-of-life care options with the patient's family.

The two clinical prediction models assess patient mortality risk by analyzing various clinical data (such as physiological and biochemical indicators and clinical scores) and constructing nomogram models. These models provide a scientific basis for medical decision-making, helping physicians to identify high-risk sepsis patients early and take appropriate treatment measures promptly. This may include enhanced monitoring, adjusting medications, or performing necessary surgical interventions, thereby reducing mortality rates and improving clinical outcomes. Early intervention is a critical factor in sepsis management, significantly increasing survival rates. The prediction models offer quantitative, data-driven risk assessments, enabling the medical team to make decisions scientifically. This not only enhances the personalization and precision of treatment but also aids in the rational allocation of medical resources in environments with limited resources.

Limitations and Future Directions

This study has certain limitations. As with all retrospective studies, there may be information bias and selection bias due to the design based on retrospective data analysis. Retrospective studies rely on existing records and data, which may lead to incomplete data. For example, medical records may lack certain key health information or fail to accurately document all clinical interventions. Additionally, the data for this study were sourced from a single center—Peking Union Medical College Hospital. Patient characteristics in different regions or medical environments may vary significantly. Although the study performed internal validation of the models, external validation has not yet been conducted. Future research should aim to include external validation and prospective validation to enhance the generalizability and practical applicability of the models. While our predictive models showed good statistical performance, their complexity might affect ease of use and interpretability in clinical practice. Therefore, clinicians using these models should have some clinical knowledge and experience to effectively integrate the prediction results with actual clinical assessments for risk evaluation and decision support.

In the future, studies could expand to multiple centers to include patient data from diverse populations and medical environments, validating the models' applicability and accuracy in different clinical contexts. Verifying model performance with datasets from various centers can improve their clinical predictive capability. When possible, constructing new clinical prediction models could include dynamic data, such as continuous monitoring and repeated measurements, to provide more real-time and accurate risk assessments^41,42. Dynamic data reflect changes in patient status over time, such as physiological parameters (heart rate, blood pressure, oxygen saturation) and repeated measurement results (e.g., glucose, electrolyte levels). These data can be collected in real-time through patient monitoring systems and automatically recorded and analyzed. Methods such as time-series analysis, dynamic Bayesian networks, or recurrent neural networks can be used to process these data, dynamically adjusting risk assessments based on the latest patient information to provide physicians with up-to-date risk insights. In clinical practice, such dynamic models can monitor changes in sepsis patients' conditions in real-time, allowing for timely adjustments to treatment plans and improving survival rates.

This study developed and validated two clinical prediction models for sepsis patients to assess in-hospital mortality and 28-day mortality post-ICU admission. Utilizing LASSO regression and logistic regression analysis, key predictive factors were identified from a multitude of clinical indicators, providing physicians with effective risk assessment tools. During model validation, both models demonstrated excellent statistical performance and clinical applicability, showing high predictive accuracy, consistency, and practicality. The stability of the models was confirmed through ROC curve, calibration curve, and decision curve analyses in both the training and validation sets. These prediction models analyze clinical and biochemical data to provide prognostic assessments for sepsis patients at different stages. They assist physicians in making informed decisions based on the patient's condition at various stages, thereby enhancing treatment effectiveness, optimizing resource allocation, and providing strong support for improving clinical outcomes in sepsis patients.

Funding Declaration

This study was supported by the National High Level Hospital Clinical Research Funding (2022-PUMCH-A-221).

Conflicts of Interest

The authors declare that there are no conflicts of interest regarding the publication of this paper. Each author has participated sufficiently in the work to take public responsibility for appropriate portions of the content and agrees that there are no relationships or associations that could be perceived as potential conflicts of interest.

Availability of Data and Materials

The datasets generated and/or analysed during the current study are not publicly available due to privacy and ethical concerns. The data may contain personal or sensitive information, and its release could violate privacy protection regulations or ethical guidelines. However, the data are available from the corresponding author on reasonable request.

Acknowledgements

None.

Singer M, Deutschman CS, Seymour CW, et al. The Third International Consensus Definitions for Sepsis and Septic Shock (Sepsis-3). JAMA. 2016;315(8):801. doi:10.1001/jama.2016.0287
Liu V, Escobar GJ, Greene JD, et al. Hospital Deaths in Patients With Sepsis From 2 Independent Cohorts. JAMA. 2014;312(1):90. doi:10.1001/jama.2014.5804
Fleischmann-Struzek C, Mellhammar L, Rose N, et al. Incidence and mortality of hospital- and ICU-treated sepsis: results from an updated and expanded systematic review and meta-analysis. Intensive Care Med. 2020;46(8):1552-1562. doi:10.1007/s00134-020-06151-x
Ebrahim GJ. Sepsis, septic shock and the systemic inflammatory response syndrome. Journal of Tropical Pediatrics. 2011;57(2):77-79. doi:10.1093/tropej/fmr022
Rudd KE, Johnson SC, Agesa KM, et al. Global, regional, and national sepsis incidence and mortality, 1990–2017: analysis for the Global Burden of Disease Study. The Lancet. 2020;395(10219):200-211. doi:10.1016/S0140-6736(19)32989-7
Shankar-Hari M, Harrison DA, Rubenfeld GD, Rowan K. Epidemiology of sepsis and septic shock in critical care units: comparison between sepsis-2 and sepsis-3 populations using a national critical care database. British Journal of Anaesthesia. 2017;119(4):626-636. doi:10.1093/bja/aex234
Huang W, Liu D, Zhang H, Ding X, Wang X, On Behalf of the Critical Care Ultrasound Study Group (CCUSG). Focus on host/organ unregulated response: a common cause of critical illness. Chinese Medical Journal. 2023;136(1):108-110. doi:10.1097/CM9.0000000000002374
Valeanu L, Bubenek-Turconi SI, Ginghina C, Balan C. Hemodynamic Monitoring in Sepsis—A Conceptual Framework of Macro- and Microcirculatory Alterations. Diagnostics. 2021;11(9):1559. doi:10.3390/diagnostics11091559
Ince C. Hemodynamic coherence and the rationale for monitoring the microcirculation. Crit Care. 2015;19(S3):S8. doi:10.1186/cc14726
Ince C, De Backer D, Mayeux PR. Microvascular Dysfunction in the Critically Ill. Critical Care Clinics. 2020;36(2):323-331. doi:10.1016/j.ccc.2019.11.003
Lesur O, Delile E, Asfar P, Radermacher P. Hemodynamic support in the early phase of septic shock: a review of challenges and unanswered questions. Ann Intensive Care. 2018;8(1):102. doi:10.1186/s13613-018-0449-8
Zanotti-Cavazzoni S, Dellinger RP. Hemodynamic optimization of sepsis-induced tissue hypoperfusion. Crit Care. 2006;10(Suppl 3):S2. doi:10.1186/cc4829
Evans L, Rhodes A, Alhazzani W, et al. Surviving sepsis campaign: international guidelines for management of sepsis and septic shock 2021. Intensive Care Med. 2021;47(11):1181-1247. doi:10.1007/s00134-021-06506-y
Lima A, Bakker J. Clinical assessment of peripheral circulation: Current Opinion in Critical Care. 2015;21(3):226-231. doi:10.1097/MCC.0000000000000194
Mouncey PR, Osborn TM, Power GS, et al. Trial of Early, Goal-Directed Resuscitation for Septic Shock. N Engl J Med. 2015;372(14):1301-1311. doi:10.1056/NEJMoa1500896
Bennett VA, Vidouris A, Cecconi M. Effects of Fluids on the Macro- and Microcirculations. Crit Care. 2018;22(1):74. doi:10.1186/s13054-018-1993-1
Wang G, Lian H, Zhang H, Wang X. Microcirculation and Mitochondria: The Critical Unit. JCM. 2023;12(20):6453. doi:10.3390/jcm12206453
for the ProCESS investigators, Massey MJ, Hou PC, et al. Microcirculatory perfusion disturbances in septic shock: results from the ProCESS trial. Crit Care. 2018;22(1):308. doi:10.1186/s13054-018-2240-5
Damiani E, Carsetti A, Casarotta E, et al. Microcirculation-guided resuscitation in sepsis: the next frontier? Front Med. 2023;10:1212321. doi:10.3389/fmed.2023.1212321
Leone M, Asfar P, Radermacher P, Vincent JL, Martin C. Optimizing mean arterial pressure in septic shock: a critical reappraisal of the literature. Crit Care. 2015;19(1):101. doi:10.1186/s13054-015-0794-z
van Genderen ME, Paauwe J, de Jonge J, et al. Clinical assessment of peripheral perfusion to predict postoperative complications after major abdominal surgery early: a prospective observational study in adults. Crit Care. 2014;18(3):R114. doi:10.1186/cc13905
Lima A, Bakker J. Clinical monitoring of peripheral perfusion: there is more to learn. Crit Care. 2014;18(1):113. doi:10.1186/cc13738
Guo Q, Lian H, Wang G, Zhang H, Wang X. Prospective Evaluation of the Peripheral Perfusion Index in Assessing the Organ Dysfunction and Prognosis of Adult Patients With Sepsis in the ICU. J Intensive Care Med. Published online May 15, 2024:08850666241252758. doi:10.1177/08850666241252758
Papageorgiou G, Grant SW, Takkenberg JJM, Mokhles MM. Statistical primer: how to deal with missing data in scientific research?†. Interactive CardioVascular and Thoracic Surgery. 2018;27(2):153-158. doi:10.1093/icvts/ivy102
Donders ART, Van Der Heijden GJMG, Stijnen T, Moons KGM. Review: A gentle introduction to imputation of missing values. Journal of Clinical Epidemiology. 2006;59(10):1087-1091. doi:10.1016/j.jclinepi.2006.01.014
Berkelmans GFN, Read SH, Gudbjörnsdottir S, et al. Population median imputation was noninferior to complex approaches for imputing missing values in cardiovascular prediction models in clinical practice. Journal of Clinical Epidemiology. 2022;145:70-80. doi:10.1016/j.jclinepi.2022.01.011
Ranstam J, Cook JA. LASSO regression. British Journal of Surgery. 2018;105(10):1348-1348. doi:10.1002/bjs.10895
Rajaratnam B, Roberts S, Sparks D, Dalal O. Lasso Regression: Estimation and Shrinkage via the Limit of Gibbs Sampling. Journal of the Royal Statistical Society Series B: Statistical Methodology. 2016;78(1):153-174. doi:10.1111/rssb.12106
Wang L, Li P, Hou M, Zhang X, Cao X, Li H. Construction of a risk prediction model for Alzheimer’s disease in the elderly population. BMC Neurol. 2021;21(1):271. doi:10.1186/s12883-021-02276-8
Huber M, Schober P, Petersen S, Luedi MM. Decision curve analysis confirms higher clinical utility of multi-domain versus single-domain prediction models in patients with open abdomen treatment for peritonitis. BMC Med Inform Decis Mak. 2023;23(1):63. doi:10.1186/s12911-023-02156-w
Zhao L, Leng Y, Hu Y, et al. Understanding decision curve analysis in clinical prediction model research. Postgraduate Medical Journal. Published online March 7, 2024:qgae027. doi:10.1093/postmj/qgae027
Piovani D, Sokou R, Tsantes AG, Vitello AS, Bonovas S. Optimizing Clinical Decision Making with Decision Curve Analysis: Insights for Clinical Investigators. Healthcare. 2023;11(16):2244. doi:10.3390/healthcare11162244
Chan JYL, Leow SMH, Bea KT, et al. Mitigating the Multicollinearity Problem and Its Machine Learning Approach: A Review. Mathematics. 2022;10(8):1283. doi:10.3390/math10081283
Wang X, Lu J, Song Z, Zhou Y, Liu T, Zhang D. From past to future: Bibliometric analysis of global research productivity on nomogram (2000–2021). Front Public Health. 2022;10:997713. doi:10.3389/fpubh.2022.997713
Iasonos A, Schrag D, Raj GV, Panageas KS. How To Build and Interpret a Nomogram for Cancer Prognosis. JCO. 2008;26(8):1364-1370. doi:10.1200/JCO.2007.12.9791
Coutrot M, Dudoignon E, Joachim J, Gayat E, Vallée F, Dépret F. Perfusion index: Physical principles, physiological meanings and clinical implications in anaesthesia and critical care. Anaesthesia Critical Care & Pain Medicine. 2021;40(6):100964. doi:10.1016/j.accpm.2021.100964
Hasanin A, Mukhtar A, Nassar H. Perfusion indices revisited. j intensive care. 2017;5(1):24. doi:10.1186/s40560-017-0220-5
Bateman RM, Sharpe MD, Ellis CG. Bench-to-bedside review: Microvascular dysfunction in sepsis –hemodynamics, oxygen transport, and nitric oxide. Crit Care. 2003;7(5):359. doi:10.1186/cc2353
Kazune S, Caica A, Volceka K, Suba O, Rubins U, Grabovskis A. Relationship of mottling score, skin microcirculatory perfusion indices and biomarkers of endothelial dysfunction in patients with septic shock: an observational study. Crit Care. 2019;23(1):311. doi:10.1186/s13054-019-2589-0
Ince C. The microcirculation is the motor of sepsis. Crit Care. 2005;9(Suppl 4):S13. doi:10.1186/cc3753
Bull LM, Lunt M, Martin GP, Hyrich K, Sergeant JC. Harnessing repeated measurements of predictor variables for clinical risk prediction: a review of existing methods. Diagn Progn Res. 2020;4(1):9. doi:10.1186/s41512-020-00078-z
Plate JDJ, Van De Leur RR, Leenen LPH, Hietbrink F, Peelen LM, Eijkemans MJC. Incorporating repeated measurements into prediction models in the critical care setting: a framework, systematic review and meta-analysis. BMC Med Res Methodol. 2019;19(1):199. doi:10.1186/s12874-019-0847-0

No competing interests reported.

SupplementaryMaterial.pdf

Download PDF

Journal Publication

published 05 Nov, 2024

Read the published version in Scientific Reports →

Editorial decision: Revision requested
08 Oct, 2024
Reviews received at journal
07 Oct, 2024
Reviewers agreed at journal
11 Sep, 2024
Reviews received at journal
29 Aug, 2024
Reviewers agreed at journal
27 Aug, 2024
Reviewers invited by journal
27 Aug, 2024
Editor assigned by journal
27 Aug, 2024
Editor invited by journal
22 Aug, 2024
Submission checks completed at journal
21 Aug, 2024
First submitted to journal
11 Aug, 2024

You are reading this latest preprint version

Construction and Validation of a Clinical Prediction Model for Sepsis Based on Peripheral Perfusion Index: In-Hospital and 28-Day Mortality Risk Prediction

Status:

Journal Publication

Version 1

Abstract

Background

Methods

Results

Conclusion

Figures

Introduction

Materials and Methods

Study Design and Setting

Study Subjects

Data Collection

Statistical Analysis

Results

Participant Characteristics

Selection of Independent Risk Factors

Development of Prediction Models

Validation of Prediction Models

Discussion

Limitations and Future Directions

Conclusion

Declarations

References

Additional Declarations

Supplementary Files

Status:

Journal Publication

Version 1