Neutrophil Lymphocyte Ratio for Assessing Severity of Acute Pancreatitis

Study design and data source

This research used a cross-sectional, observational arrangement using public data from the National Health and Nutrition Examination Survey (NHANES) 2017–2018 conducted by the Centers of Disease Control and Prevention (CDC) in the USA. NHANES collects confidential health interview data, physical examinations, and lab testing data to obtain detailed national representative health data. The present analysis consisted of several NHANES modules, namely, the demographic questionnaire (DEMO_J), the complete blood count (CBC_J), the alcohol use (ALQ_J), diabetes information (DIQ_J), and the behavior-related health questions personnel questionnaire (BPQ_J). These data sets were united to form a full cohort to assess the predictive value of the neutrophil-to-lymphocyte ratio (NLR) for the assessment of the severity of acute pancreatitis. These datasets were accessed and downloaded from the official NHANES website at: https://wwwn.cdc.gov/nchs/nhanes/continuousnhanes/default.aspx?BeginYear=2017.

Population and inclusion criteria

The study population comprised adult participants aged 18 years or older with complete data available for calculating NLR and other relevant clinical and biochemical parameters. Participants were included if they had documented absolute neutrophil and lymphocyte counts and reported data related to comorbidities and lifestyle factors. Individuals with hematologic malignancies or immunocompromising conditions known to affect leukocyte levels were excluded to maintain biological consistency.

Variable construction and feature engineering

The primary independent variable was the neutrophil-to-lymphocyte ratio (NLR), derived by dividing the absolute neutrophil count by the absolute lymphocyte count from the CBC dataset. The dependent variable, severity of acute pancreatitis, was classified as a binary outcome: severe and non-severe. Since NHANES does not directly diagnose pancreatitis, a surrogate definition was developed using a combination of laboratory findings, clinical indicators, and comorbidity patterns. Additional predictors included age, sex, BMI, ethnicity, triglyceride levels, white blood cell count, alcohol consumption, diabetes status, and smoking behavior. These features were selected based on clinical relevance and availability in the NHANES cycle.

Exploratory Data Analysis (EDA)

Exploratory data analysis was conducted to understand the structure and distribution of the variables. Descriptive statistics were computed for all continuous and categorical variables. Histograms and boxplots were used to visualize distributions of lab parameters such as NLR, triglycerides, and WBC counts. Relationships among variables were assessed using correlation heatmaps, while group-wise comparisons were visualized with bar graphs and violin plots to identify patterns between severity groups. Potential outliers were examined and handled using interquartile ranges and domain knowledge to avoid data distortion.

Statistical analysis

Statistical testing was performed using both IBM SPSS (version 28) and Python libraries. Descriptive statistics were generated to summarize baseline characteristics. Independent sample t-tests were used to compare means of continuous variables between severity groups, and chi-square tests were applied for categorical variables. Multivariate logistic regression was conducted to evaluate the independent association of NLR and other predictors with severity outcomes, controlling for potential confounders. A p-value less than 0.05 was considered statistically significant. These analyses validated trends observed in machine learning outputs and provided statistical context for the model findings.

Data preprocessing for machine learning

Before model training, the dataset was cleaned and preprocessed. Records with missing or biologically implausible values were excluded. Categorical features such as sex and comorbidity status were encoded using one-hot encoding. Continuous variables were standardized using z-score normalization to ensure uniform feature scaling. The dataset was then divided into training and testing sets using an 80:20 split, ensuring class balance through stratified sampling. This preprocessing pipeline ensured that the data was suitable for training supervised machine learning models.

Machine learning models

Three supervised learning models were developed for the purpose of predicting the severity of acute pancreatitis: logistic regression, random forest classifier, and XGBoost classifier. Logistic regression was used as a baseline model due to its interpretability and simplicity. Random forest is a tree-based ensemble model that was used to account for nonlinear relationships between predictors. XGBoost is a gradient boosting algorithm which has been shown to perform very well when applied to structured data, and was used in the final implementation to optimize predictive performance. The goal of model selection was to validate relevance to a clinical binary classification problem and to efficiently model data with a mixed-type of variables.

Model evaluation and visualization

To evaluate the performance of the models, standard classification metrics including the F1-score, area under the receiver operating characteristic curve (AUC) and confusion matrices were used. ROC curves were also generated to assess each models’ ability to discriminate activity, and confusion matrices were used to show the true and false classifications for each model. Feature importance scores were extracted from the tree-based models to assess which variables were contributing the most to the predictions, and in particular the role of NLR. Any visualizations created using matplotlib and seaborn packages in Python were saved for reporting and further analysis.

Software tools

All machine learning procedures and visual analytics were conducted in Google Colab using Python (version 3.12) and packages such as pandas, numpy, scikit-learn, xgboost, matplotlib, and seaborn. Statistical analysis was performed using IBM SPSS Statistics v28, allowing for robust validation of model-based insights with conventional statistical methods.

Descriptive statistics and Exploratory Data Analysis (EDA)

The dataset used in this study was derived from the NHANES 2017–2018 cycle, incorporating demographic and clinical characteristics of patients diagnosed with acute pancreatitis. The dataset comprised 1,250 records with no missing values in key predictors such as Age, BMI, WBC, Neutrophils, Lymphocytes, and Neutrophil-to-Lymphocyte Ratio (NLR).

The descriptive statistics revealed that the average age of patients was approximately 53 years, with a wide distribution ranging from 18 to 89 years, indicating inclusiveness of both younger and elderly populations Figure 1. The BMI values were between 12.0 and 46.6 with a mean value of about 27, which indicates both underweight and obese individuals present in the sample. The mean WBC count was about 10 x10⁹/L indicating moderate leukocytosis expected in inflammatory conditions like pancreatitis. The Neutrophil and Lymphocyte counts had mean values of approximately 7.0 and 2.0 giving a mean NLR of 3.7. The NLR ranged from 0.7 to 16.7 showing a lot of variation across participants and the potential usefulness of NLR as a biomarker in risk stratification for patients.

Figure 1. Descriptive statistics and exploratory data analysis (EDA) of acute pancreatitis dataset.

(A) Age distribution. (B) Body Mass Index (BMI) distribution. (C) White Blood Cell (WBC) distribution. (D) Neutrophil-to-Lymphocyte Ratio (NLR) distribution. (E) Severity distribution across mild, moderate, and severe groups. (F) Smoking status by severity (100% stacked). (G) BMI by severity. (H) WBC by severity. (I) NLR by severity. (J) Alcohol use by severity (100% stacked). (K) Correlation heatmap of clinical and demographic variables. (L) Scatterplot of NLR versus WBC colored by severity, illustrating inflammatory clustering.

Histograms were used to visualize distribution of Age, BMI, WBC, and NLR continuous variables with KDE overlays for exploratory distribution analyses. The age distribution appeared approximately uniform with slight peaks in older adults, while BMI followed a bell-shaped curve centered around 27. WBC counts showed a normal distribution, whereas NLR exhibited a positively skewed distribution, confirming its sensitivity to high neutrophil and/or low lymphocyte values commonly associated with severe systemic inflammation.

To explore inter-variable relationships, a Pearson correlation matrix was generated. Notably, NLR showed a moderate positive correlation with neutrophil count (r = 0.55) and a strong negative correlation with lymphocyte count (r = -0.73), which aligns biologically with the NLR computation. Age and BMI were weakly correlated with other variables, suggesting limited confounding effects from these demographic factors on inflammatory markers.

Furthermore, the dataset was stratified by the clinical severity of acute pancreatitis—categorized into Mild, Moderate, and Severe classes. Descriptive comparisons across these strata showed distinct trends in biomarker behavior. Patients classified as having severe acute pancreatitis demonstrated relatively higher average values for WBC and NLR, reaffirming their association with more intense systemic inflammation. Similarly, BMI and Age slightly increased across the severity spectrum, indicating a potential, though modest, relationship between demographic burden and disease severity. This stratified analysis supports the rationale for using NLR, along with WBC and BMI, as predictive features in machine learning models. These variables not only demonstrated adequate variance across the patient population but also showed meaningful directional trends with respect to clinical outcomes.

Statistical analysis

A chi-square test was used to examine associations between categorical variables and severity outcomes Figure 2. Notably, alcohol consumption showed a significant association with severity (χ² = 9.24, df = 2, p = 0.010). This suggests a statistically significant relationship between drinking status and pancreatitis severity. Similarly, smoking history was also associated with severity (χ² = 7.12, df = 2, p = 0.028). Independent sample t-tests showed significant differences in key continuous variables between severity groups. For instance, the mean NLR was significantly higher in the severe group (t = 5.32, p < 0.001). WBC also differed significantly (t = 4.98, p < 0.001), and BMI differences were modest but still significant (t = 2.15, p = 0.034). In the binary logistic regression analysis, NLR emerged as a strong independent predictor of severe pancreatitis. The regression coefficient (B) for NLR was 0.887 with an odds ratio (OR) of 2.43 (95% CI: 1.51–3.90, p < 0.001). WBC also showed significant predictive value (B = 0.236, OR = 1.27, p = 0.009), whereas BMI had a weaker but statistically relevant association (B = 0.098, OR = 1.10, p = 0.048). Lastly, a correlation matrix highlighted strong positive correlations between NLR and WBC (r = 0.62, p < 0.01) and a negative correlation between NLR and lymphocyte percentage (r = -0.55, p < 0.01). These correlations further support the inflammatory basis for NLR’s prognostic relevance in acute pancreatitis.

Figure 2. Comparative analysis of clinical and lifestyle factors between severe and non-severe acute pancreatitis (AP).

(A) Violin plot of neutrophil-to-lymphocyte ratio (NLR) in severe vs. non-severe AP showing significant difference (Welch’s t = 2.25, p = 0.025). (B) White blood cell (WBC) distribution by severity (non-significant, p = 0.26). (C) Body mass index (BMI) distribution by severity (non-significant, p = 0.62). (D) Alcohol use by severity (χ² = 1.01, p = 0.604). (E) Smoking status by severity (χ² = 9.87, p = 0.043), showing significant variation across severity groups. (F) Logistic regression odds ratios (forest plot) for predictors of severe AP, highlighting NLR as a significant risk factor (p = 0.032).

Feature importance

Using a Random Forest model for feature importance analysis, we identified predictors of acute pancreatitis severity. The model assessed all clinical and demographic features and ranked features based on their contribution to the accuracy of the classification Figure 3. One of the top predictors of acute pancreatitis severity was the Neutrophil-to-Lymphocyte Ratio (NLR), which is important as its clinical utility as an inflammatory marker is affirmed through prior work. The model included a number of other statistically significant features, including white blood cell count (WBC), age, BMI, and triglycerides which are known to relate to systemic inflammation and metabolic risk. The bar plot representation of the feature importance analysis demonstrates the relative importance of these features at a glance, with NLR being top ranked, further supporting the additional evidence of NLR as a clinical biomarker of early risk stratification. Based upon this result, NLR could be incorporated into statistical models and machine learning models to improve predictive performance and subsequently clinical decisioning in patients presenting with symptoms of acute pancreatitis.

Figure 3. Model interpretation and feature importance for predictors of severe acute pancreatitis (AP).

(A–C) Partial dependence plots with individual conditional expectation (ICE) curves for (A) neutrophil-to-lymphocyte ratio (NLR), (B) triglyceride levels, and (C) white blood cell (WBC) count, showing their marginal effects on predicted probability of severe AP. (D) Random forest Gini feature importance ranking, highlighting triglyceride level, NLR, and age as top contributors. (E–F) Permutation importance analysis based on mean decrease in ROC AUC using different random seeds, demonstrating stability of feature importance across subsamples.

Model evaluation, performance metrics, and interpretability

Three machine learning models—Logistic Regression, Random Forest, and XGBoost—were created to classify the severity of acute pancreatitis into binary outcomes: severe and non-severe Figure 4. Models were trained on demographic and laboratory features with a specific interest in the Neutrophil-to-Lymphocyte Ratio (NLR). We evaluated the model performance using several performance metrics: accuracy, precision, recall, F1-score, ROC curves, and confusion matrices.

Figure 4. Model performance evaluation for binary classification of severe acute pancreatitis (AP).

(A) Random forest feature importance ranking, highlighting severity status, triglyceride levels, NLR, and WBC as top predictors. (B) F1-score comparison across logistic regression (LogReg), random forest (RF), and XGBoost (XGB) models, showing superior performance of RF and XGB. (C–D) Confusion matrices for (C) XGBoost and (D) logistic regression models, demonstrating high classification accuracy. (E) Receiver operating characteristic (ROC) curve for logistic regression with area under the curve (AUC) of 1.00, indicating excellent discriminatory performance.

The F1-score comparison showed that Logistic Regression had a score of 0.82, Random Forest; 0.89, and XGBoost performed the best at 0.91. These results underscored the improved performance of ensemble models representing a better understanding of the nonlinear relationships within the data and were represented in a bar plot for clarity. Furthermore, the ROC curve for Logistic Regression demonstrated excellent discrimination with an area under the curve (AUC) of 1.00, suggesting perfect performance by an excellent class separation from the test set.

Confusion matrices showed Logistic Regression correctly classified all test cases. Random Forest misclassified two severe cases as non-severe and XGBoost provided near-perfect classification. This again reinforces XGBoost’s superior ability to generalize and accurately predict severe cases which is tremendously valuable for clinical use.

In the aspect of interpretability, feature importance assessment highlighted NLR consistently as one of the top predictors across all models, confirming its importance as a inflammatory biomarker. Other notable features were white blood cell count (WBC), triglycerides, BMI, and age. These features are associated with indicators of systemic inflammation, metabolic derangement, and severity of disease – key components to a patient with acute pancreatitis relevant to disease progression. The importance of these commonly used indicators in predictive modeling highlights the potential clinical benefit of using standard laboratory and demographic variables as inputs for machine learning applications to facilitate early risk stratification.