Unsupervised Machine Learning Drives Dynamic Alignment Classification in Navigated Total Knee Arthroplasty

Current classifications used for total knee arthroplasty (TKA) are static and fail to capture the dynamic behavior of the limb during gait. This study introduces a novel intraoperative method to measure dynamic hip-knee-ankle (dHKA) angle using an intra-articular device coupled with a computer-assisted orthopaedic surgery system. This device applies a quasi-constant distraction force throughout the knee joint range of motion. A machine learning model was utilized to identify natural data groupings and develop a classification based on patient-specific dHKA profiles. We analyzed dHKA before and after the femoral cut (tibia-first TKA workflow) and assessed how often postcut clusters matched precut clusters across surgeons.

A total of 1,890 tibia-first TKA cases performed by 11 surgeons were reviewed. For each case, HKA angles were recorded at 12 flexion angles (0 to 120°) before and after the femoral cut. Using precut data, a 12-dimensional map was created with each dimension representing the degree of HKA at a specific flexion angle. A K-means clustering model was trained on data collected before the tibial cut to identify alignment profiles. The trained model was then applied to data collected after the tibial cut for comparison. A subset of 141 TKA cases from a single surgeon who had one-year Knee Injury and Osteoarthritis Outcome Score for Joint Replacement scores was analyzed to explore the association between cluster preservation and early functional outcomes.

Clustering evaluation identified four clusters and an eight-dimensional feature space as optimal. Precut/postcut cluster distributions were cluster 1 (15.3%/14.2%), cluster 2 (35.9%/30.1%), cluster 3 (34.2%/35.6%), and cluster 4 (14.5%/20.2%). Cluster 1 was characterized as valgus and neutral, cluster 2 as neutral, cluster 3 as low-to-moderate varus, and cluster 4 as moderate-to-high varus. Overall, 69.4% of the cases retained the same cluster postcut, with surgeon-specific match rates ranging from 61 to 88%. In the outcomes subset, 72.3% preserved their precut cluster. Preservation was associated with greater Knee Injury and Osteoarthritis Outcome Score for Joint Replacement improvement, with cluster-specific significance observed in low-to-moderate varus knees.

This study demonstrates the first use of unsupervised machine learning to classify intraoperative dHKA profiles captured with a force-controlled intra-articular device and computer-assisted orthopaedic surgery system. This enables real-time feedback and offers a foundation for an automated alignment classification guidance in personalized TKA.