Diagnostic value of integrated albumin-to-leukocyte ratio risk modeling stratification and machine learning approaches for prediction of spontaneous bacterial peritonitis in cirrhotic patients.

backgroundSpontaneous bacterial peritonitis (SBP) remains a life-threatening complication of liver cirrhosis, requiring accurate and rapid prediction. Albumin and leukocyte count reflect systemic inflammation and liver dysfunction; combining them into an Albumin-to-Leukocyte Ratio (ALR) may enhance early risk stratification.

aimTo evaluate the diagnostic value of the ALR for predicting SBP in patients with cirrhosis and develop an ALR-based risk score. We will compare its predictive accuracy with machine-learning models to establish a practical triage tool for SBP risk stratification. PATIENT AND

methodsThis cross-sectional study included 240 cirrhotic patients, categorized into SBP (n = 152) and non-SBP (n = 88). Demographic, clinical, hematological, and biochemical variables were analyzed. ALR was calculated, and its predictive accuracy was assessed using ROC curves. A risk score integrating ALR, hemoglobin, bilirubin, and Child-Pugh class was generated. Logistic regression identified independent predictors. Diagnostic performance of ALR was compared with decision tree and random forest models.

resultsSBP patients had significantly lower albumin, ALR, hemoglobin, and platelets, and higher leukocyte count, bilirubin, INR, and creatinine (all p ≤ 0.05). ALR demonstrated moderate diagnostic performance with fair discrimination; AUC of 0.707, achieving a sensitivity of 82% at a cutoff of ≤ 0.728. Multivariate analysis identified ALR (OR = 0.114, p = 0.001), hemoglobin (OR = 0.847, p = 0.024), bilirubin (OR = 1.319, p = 0.004), MELD score (OR = 0.998, p = 0.042) and Child-Pugh class C (OR = 4.86, p = 0.001) as independent predictors. The generated ALR risk score (0-7 points) categorized patients into low, intermediate, and high-risk groups with increasing diagnostic accuracy (AUC range 0.63-0.75). Machine-learning models outperformed ALR alone; random forest showed the highest discrimination (AUC = 0.901), followed by decision tree (AUC = 0.852).

conclusionALR is a simple and accessible biomarker with moderate accuracy for predicting SBP in cirrhotic patients and may serve as a practical bedside triage tool when integrated into a composite risk score. However, machine-learning models, particularly random forest, demonstrated superior predictive performance, suggesting that combining conventional biomarkers with advanced algorithms may enhance SBP risk stratification. CLINICAL TRIAL NUMBER: Not applicable.