A Nearest Neighbors Classifier For Automatically Determining Foot Strike Patterns In Runners

The foot angle at initial contact (foot strike angle) is often used to distinguish between forefoot (FFS), midfoot (MFS), and rearfoot (RFS) strike runners. However, there is little consensus on which cut-points should be used for classification, leading to questions about generalizability. PURPOSE: To develop a machine learning classifier for determining foot strike patterns of runners and compare its performance to methods relying on measuring foot strike angle and using discrete cut-points for classification. METHODS: Sagittal plane video recordings of 724 U.S. Army trainees undergoing a treadmill running assessment in an ongoing prospective study were used for this study. Coordinates of the heel and 5th metatarsal head were extracted from these recordings using DeepLabCut and used to compute foot strike angles. A weighted K nearest neighbors (KNN) classifier was trained to distinguish FFS, MFS, and RFS using 80% of the data. The remaining 20% of the data were used to assess the accuracy (true-positive rate) of the KNN classifier. Foot strike patterns were then determined from foot strike angles using the 3 most frequently cited cut-points. Reference data was obtained from visual inspection of the videos (gold standard). RESULTS: The KNN classifier had equal or greater accuracy for FFS, RFS, and overall compared to cut-point methods. Two cut-point methods had greater MFS accuracy, but considerably lower overall and RFS accuracy. CONCLUSION: The KNN classifier was more accurate than the simple cut-point methods for determining foot strike patterns, generalizing well to the data set. This classifier provides researchers and practitioners with a generalizable tool for classifying the foot strike patterns of runners. Supported by US DoD award W81XWH-20-C-0123 P0001. Opinions in this manuscript are the private views of the authors, not to be construed as official policy or reflecting the views of the US Army. Table 1. - Accuracy (true-positive rates) for overall, forefoot (FFS), midfoot (MFS), and rearfoot (RF Method Overall Forefoot Midfoot Rearfoot (n = 145) (n = 12) (n = 6) (n = 127) KNN 91.7% 91.7% 33.3% 94.5% Lieberman et al.a 87.6% 91.7% 33.3% 89.8% Forrester & Townendb 67.6% 25.0% 100.0% 70.1% Altman & Davisc 86.2% 91.7% 100.0% 85.0% Supported by US DoD award W81XWH-20-C-0123 P0001. Opinions in this manuscript are the private views of the authors, not to be construed as official policy or reflecting the views of the US Army. AI was for data processing and analysis in this study.