COMPARATIVE EVALUATION OF LRINEC, J-LRINEC, AND MACHINE LEARNING-BASED LRINEC MODELS FOR DIFFERENTIATING NECROTIZING FASCIITIS FROM SEVERE CELLULITIS

Abstract

Background: Differentiating early-stage necrotizing fasciitis (NF) from severe cellulitis is a critical clinical challenge in acute care settings. While the Laboratory Risk Indicator for Necrotizing Fasciitis (LRINEC) score is widely used, real-world validation studies report highly variable diagnostic performance across different populations and pathogen profiles. This study aimed to evaluate and compare the diagnostic and prognostic accuracy of the traditional LRINEC score, a regionally recalibrated J-LRINEC score, and an internally developed machine learning-modified LRINEC (MLRINEC) model to optimize early detection and intervention protocols. Methods: A retrospective single-centre comparative observational study was conducted at the Department of General Surgery, Saveetha Medical College and Hospitals, Chennai, India. Consecutive adult patients presenting to the emergency department (ED) with suspected severe soft tissue infections between January 2023 and December 2025 were screened from emergency department admission registers, surgical operation logs, and histopathology records. A total of 500 adult patients (100 confirmed NF cases and 400 severe cellulitis controls) met the eligibility criteria after a documented screening and exclusion process. Clinical and laboratory parameters drawn within 2 hours of ED triage were extracted. Diagnostic performance was assessed using Receiver Operating Characteristic (ROC) curves with 95% confidence intervals, and independent clinical predictors were identified via multivariate logistic regression analysis. The MLRINEC model was developed and internally validated using an 80/20 train-test split with 5-fold cross-validation; it has not undergone external or prospective validation. Results: At the traditional cutoff of 6 or above, the original LRINEC score demonstrated a sensitivity of 68.0% (95% CI: 58.4–77.6) and a specificity of 85.0% (95% CI: 81.5–88.5), with an AUC of 0.785 (95% CI: 0.732 0.838). The J-LRINEC score, applied using its published Japanese-derived coefficients without population specific re-derivation rationale and limitation discussed in Section IV, achieved a sensitivity of 88.0% (95% CI: 81.2–94.8) and specificity of 81.0% (95% CI: 77.1–84.9), with an AUC of 0.924 (95% CI: 0.886–0.962). An internally developed Random Forest MLRINEC model showed an AUC of 0.968 (95% CI: 0.938–0.998) on the internal 20% validation split, with a sensitivity of 92.0% (95% CI: 86.2–97.8) and specificity of 89.5% (95% CI: 86.4–92.6); these figures reflect internal validation only and require external prospective confirmation before any claim of clinical reliability. On feature importance analysis, haemoglobin levels and white blood cell count emerged as the most influential predictors of fascial necrosis in this model. Traditional scoring performed poorly in patients with diabetes (specificity: 58.0%) and in the small exploratory Vibrio vulnificus subgroup (n=12; sensitivity 35.0%). Conclusion: The traditional LRINEC score is limited by suboptimal sensitivity for early-stage triage, frequently missing the critical window for intervention. The recalibrated J-LRINEC score and an internally developed Random Forest MLRINEC model showed improved diagnostic sensitivity and specificity in this cohort; however, the MLRINEC model requires prospective external validation before it can be considered a generalisable clinical tool. Both calibrated approaches may assist clinicians as adjuncts to, not replacements for, clinical judgement in early diagnosis and surgical decision-making.