Using Machine Learning to Identify the Optimal Limb Symmetry Index Cut-Off Threshold in Paediatric Patients with ACL Injury

Abstract

Background:The risk for a subsequent ACL injury following a primary paediatric ACL reconstruction is extremely high, with 17-30% of paediatric patients sustaining a second ACL injury within two years. It is possible that this risk is associated with a lack of evidence for identifying return-to-activity (RTA) measures to guide postoperative decision-making. Limb symmetry indices (LSI) derived from functional and strength tasks are commonly used, however LSIs used by paediatric orthopaedic surgeons can range from 75-95%, with no objective evidence to support these values.Purpose:To create knowledge that could be used to inform RTA and to determine if a machine learning approach could objectively identify LSI thresholds from functional tasks that classify paediatric individuals as injured or healthy.Methods:Forty-two patients (30 females) who had suffered an ACL injury (ACLi) and 69 matched uninjured controls (36 females; CON) performed isometric knee extension and flexion, and single-leg hopping tasks. LSIs were calculated as injured/uninjured limb for ACLi and non-dominant/dominant limb for CON. LSIs with significant between-group (ACLi vs CON) differences (independent t-tests), were used in a machine learning algorithm. A classification tree (CT) was used to classify ACLi and CON participants using LSI percentages. Ten-fold cross-validation was used to determine optimal tree complexity. LSIs were then converted from percentage to categorical ‘pass/fail’ grades according to cut-off thresholds ranging from 70-100%. CTs were then fit to the converted categorical data at each cut-off threshold and the model accuracies were compared.Results:Anterior, lateral, triple, and timed 6m hop tasks, and knee extension strength LSIs were used in the machine learning algorithm. The CT using the raw LSI percentages correctly classified 85.9% of participants. When data were converted from percentage to categorical ‘pass/fail’ grades, a plateau in classification accuracy occurred between LSIs cut-off thresholds of ˜83-90%. The most accurate LSI cut-off threshold was 89%, which correctly classified 84.4% of participants.Conclusion:Based on our preliminary analysis, the 90% LSI cut-off is an appropriate threshold for separating the hopping and strength performance of ACLi and CON paediatric participants. It is nevertheless important to recognize that this threshold can identify those with ACL injury, and not those who are ready for return to activity. A 90% LSI is therefore only a minimum criterion to reach as part of an RTA evaluation, and not a threshold that identifies when an individual is ready to return.