TPW 5.01 Optimization of Emergency Department Triage in Acute Diverticulitis Through a Guideline-Integrated Machine Learning–Based Clinical Decision Support System for Hospital Admission Decisions
TL;DR
This model provides a reliable clinical “second opinion” by integrating structured biochemical data with NLP-extracted risk factors and demonstrates significant potential for enhancing triage accuracy and ensuring guideline-compliant management of acute diverticulitis.
Abstract
Standardizing emergency admission decisions for acute diverticulitis (AD) is essential to optimize bed utilization and patient safety. This study evaluates a AI based machine learning (ML)–based model designed to categorize AD severity and recommend admission versus outpatient management by integrating structured and free-text data. Data of all patients admitted with AD in specified time period was analysed using a hybrid natural language processing (NLP) and ML framework. The model utilized 14 key parameters, including demographics (age), physiology (HR, SBP, Temp, GCS), and biochemistry (CRP, WCC, Creatinine, INR). Regex-based NLP extracted clinical features from triage notes, and CT reports, specifically identifying LIF pain, and immunosuppression. Severity grading and admission logic were aligned with NICE, ACPGBI, and WSES guidelines, incorporating specific thresholds such as CRP>150 (warning level) and >200 mg/L (emergency level) to identify complicated disease. A logistic regression classifier with balanced weights was trained and evaluated using an 80/20 train-test split. The model successfully stratified patients into three severity grades. Grade 3 (Severe) was triggered by organ dysfunction or complications (perforation/abscess), while Grade 2 (Moderate) identified high-risk factors like age>65, ASA⩾3, or immunosuppression. The ML classifier demonstrated high predictive performance, achieving test-set ROC AUC of 1.0, indicating perfect discrimination between the rule-based admission recommendations and clinical features within this pilot cohort. By integrating structured biochemical data with NLP-extracted risk factors, this model provides a reliable clinical “second opinion.” It demonstrates significant potential for enhancing triage accuracy and ensuring guideline-compliant management of acute diverticulitis.