Artificial Intelligence (AI) in a Singaporean Emergency Department: Detecting fractures and reducing recalls

Abstract

Introduction: There has been rapid increase in the number of artificial intelligence and machine learning (ML) algorithms in recent years. In our local emergency department (ED), after-hours, radiographs are read by the ED doctor, with formal reporting by the radiology department performed on the subsequent day. Discrepant diagnoses between the ED doctor and radiologist potentially result in recalls of discharged patients for additional treatment, leading to greater monetary and manpower costs. To the authors' knowledge, no Singapore based study has utilized local data to analyse the performance of an AI fracture detection solution in the Singapore ED. The objective of this study is to evaluate the diagnostic performance of an AI radiograph fracture tool compared to ED doctors.

Materials and methods: A retrospective study was conducted on 42 discrepant radiographic studies. In these studies, the final radiology report by the radiology department (the "ground truth") had a different diagnosis from bedside radiographic assessment by an ED Doctor.

Results: There were 20 studies with fractures and 22 studies with no fractures. The AI solution correctly diagnosed 15 fractures (75.0% of cases with fracture) (Figure 1), missed 5 fractures (25.0% of cases with fracture) and overcalled 1 fracture (4.5% of cases with no fracture) (Figure 2). The AI solution sensitivity is 75.0%, specificity is 95.5%, positive predictive value (PPV) is 93.8% and the negative predictive value (NPV) is 80.8%.

Conclusion: Having a fracture detection AI solution has the potential of reducing discrepant cases by up to 73.7% in the ED setting. Further large-scale studies should be performed to quantify the economic, manpower and healthcare outcome benefits of such an AI solution.