Validation d’un système d’aide à la décision de l’indication chirurgicale basé sur l’IA pour le triage diagnostique de la douleur abdominale aiguë

Abstract

Introduction:Acute abdominal pain (AAP) accounts for 5–10% of emergency department visits and represents a major diagnostic challenge due to the wide range of possible aetiologies, from self-limiting benign conditions to life-threatening surgical emergencies [1,3]. Delayed diagnosis of a surgical emergency is associated with significantly increased morbidity and mortality, particularly in elderly or immunocompromised patients [7,15]. Artificial intelligence (AI) has expanded considerably in medicine, notably through Clinical Decision Support Systems (CDSS) [35]. Machine learning models trained on clinical, laboratory, and imaging data have shown promising performance for aetiological diagnosis and prediction of the need for surgery in acute abdominal pain, with areas under the curve ranging from 0.85 to 0.95 [32,39]. However, most of these models have not been validated in real-world settings, their explainability remains limited, and their clinical integration faces technical and regulatory barriers [27,42,43]. Objective:To develop and validate an artificial intelligence-based application for determining the need for surgical intervention in acute abdominal pain. Methods:A total of 205 medical records of acute abdominal pain cases were collected at the Hôpital Mixte de Laghouat and the Centre Hospitalier Universitaire Mustapha Bacha, from 1 November 2025 to 12 January 2026, in accordance with the operative model “Appendix I”. These cases were then sorted and all records that did not meet the predefined inclusion criteria for the study were excluded. Ultimately, 87 medical records were deemed eligible for analysis. The data from these records were submitted to the previously designed and programmed Kashef AI model. The sensitivity, specificity, PPV and NPV of the model for diagnosing the degree of urgency of abdominal pain were then evaluated. Results :Eighty-seven patients were recruited and tested with the Kashif AI model. An acceptable sensitivity (75%) and an excellent negative predictive value (90%) were found, potentially making it a useful primary triage tool to identify patients at low risk of surgical emergency. In contrast, its poor specificity (50.7%) and very low positive predictive value (25.5%) formally contraindicate its use as a diagnostic confirmation tool. Conclusion:The developed AI application constitutes a promising decision-support tool for surgical indication in acute abdominal pain, capable of improving diagnostic speed and accuracy in the emergency department.