Development of a Machine Learning Model for Determining Alignment in Knees Following Total Knee Arthroplasty

Malalignment remains a major reason for implant failure following total knee arthroplasty (TKA). Manual analysis of images at a large scale is untenable, and machine learning (ML) models may be a useful tool for determining alignment following TKA and can help identify patients who are at risk for failure. We aimed to develop an ML model that can accurately determine TKA alignment from full-length hip-to-ankle films.

Long-leg radiographs from 550 patients who had undergone primary TKA were collected, of which 440 were utilized to train an ML model to identify landmarks used for alignment measurement. These landmarks included the femoral head, the intertrochanteric region, the femoral midshaft, the distal aspects of the medial and lateral implant femoral condyle, the medial third and lateral third of the superior border of the tibial base plate, the tibial midshaft, and the tibial plafond. Following identification of these landmarks, mechanical hip-knee-ankle angle (mHKA), lateral distal femoral angle (LDFA), medial proximal tibia angle (MPTA), and joint line obliquity (JLO) were calculated. A separate random sample of 110 long-leg radiographs was utilized to test the accuracy of the ML model.

The ML model was able to analyze a single image in under 0.1 seconds. The mean angle error between human-measured and ML model-measured mHKA was 0.08° (SD = 0.8), the mean error for LDFA was 0.7° (SD = 0.9), the mean error for MPTA was 0.4° (SD = 0.3), and the mean error for JLO was 0.7° (SD = 0.8).

We describe an ML model that demonstrates high accuracy in determining alignment following TKA and demonstrates the vast potential of applying ML technology to improving clinical workflow and increasing research productivity in total joint arthroplasty.