2025 ISAKOS Congress in Munich, Germany

2025 ISAKOS Biennial Congress Paper


An Artificial Intelligence Based Approach to Musculoskeletal Acute Knee Injury Triage

Shehzaad Aziz Khan, MD, FRCS (Tr&Orth), North Weald, Essex UNITED KINGDOM
Dylan Mistry, MD, MRCS, Solihull, West Midlands UNITED KINGDOM
Alastair Stephens, MD, MRCS, Warwick UNITED KINGDOM
James Dalrymple, MD, FRCS (Tr&Orth), London UNITED KINGDOM

Frimley Health NHS Trust, Windsor, UNITED KINGDOM

FDA Status Not Applicable

Summary

We have shown a proof of concept in using an AI based machine learning algorithm to identify patients with acute knee injuries who require an MRI.

Abstract

Intro
In 2023, NHS England reported 230,000 of accident and emergency attendances due to acute knee injuries with an average waiting time of 16.5 weeks to see a specialist for diagnosis. We know these injuries are a significant financial and resourceful burden on our healthcare system. There has been a huge drive and investment in artificial intelligence (AI) over the last decade.

We aim to present a proof of concept study to assess if using AI technology to triage acute knee injuries can correctly identify patients with knee injuries who require Magnetic Resonance Imaging (MRI) prior to specialist review, ultimately streamlining the triage process.

Methods

A retrospective review was performed of patients in a single centre who had attended an acute knee clinic (AKC) and been reviewed by a senior orthopaedic surgeon. A 17 point questionnaire was designed and inputted in to an algorithm using rule-based expert system AI. This emulates the decision-making ability of a human expert, which in our case, is a consultant knee surgeon. All questions were written in a format that a lay person would be able to answer without the need of a medical examination. Answers were multiple choice or binary options. The symptoms and signs from the clinic letters were inputted in to the AI algorithm questionnaire to produce a result if the patient required an MRI scan and a provisional diagnosis. Primary outcome measure was correct identification of patients by algorithm who needed an MRI scan. Secondary outcome was the ability to make the correct diagnosis.

Results

201 patients were included (130 male, 71 female); average age 29.1 (16-39). 198 patients went on to have an MRI scan performed as a result of the AKC consultation. Injury diagnosis from the AKC consultation were 122 ligament, 52 menisci, 12 minor fracture (osteochondral), 5 normal, 6 patellofemoral and 4 pending results. Our algorithm correctly identified 92% of patients (185) who needed a scan and identified the correct diagnosis in 81% of patients (163). 59 patients had dual diagnosis (i.e. ligament and menisci injury) and the AI algorithm was able to correctly identify both diagnosis in 63% of patients (38). 5 patients (2.5%) were not recommended a scan from the algorithm. Further analysis showed that two patients had undisplaced posterior horn meniscal tears in an age group which would not routinely require initial surgical treatment and the other three had grade 1 collateral ligament sprains. 4 patient's (1.9%) did not have a scan, which the algorithm triaged as a ligament injury but were actually sprains on clinical assessment.

Conclusion

We have shown that an AI based algorithm can correctly identify patients with acute knee injuries presenting to primary/secondary care that require MRI imaging with a very high sensitivity and low specificity. This could improve the efficiency and cost effectiveness of the normal patient pathway for these injuries. Ultimately resulting in quicker diagnosis and treatment of these injuries.