Supervised Machine Learning for Aiding Diagnosis of Knee Osteoarthritis:
A Systematic Review and Meta-Analysis
Abstract
Background
Knee osteoarthritis (OA) remains a leading aetiology of disability
worldwide. Clinical assessment of such knee-related conditions has
improved with recent advances in gait analysis. Despite being a gold
standard method, gait data acquired by motion capture (mocap) technology
are highly non-linear and dimensional, which make traditional gait
analysis challenging. Thus, extrinsic algorithms need to be used to make
sense of gait data. Supervised Machine Learning (ML)-based classifiers
outperform conventional statistical methods in revealing intrinsic
patterns that can discern gait abnormalities when using mocap data,
making them a suitable tool for aiding diagnosis of knee OA.
Research question
Studies have demonstrated the accuracy of supervised ML-based
classifiers in gait analysis. However, these techniques have not gained
wide acceptance amongst biomechanists for two reasons: the reliability
of such methods has not been assessed and there is no consensus on which
classifier or group of classifiers to select. Specifically, it is not
clear whether classifiers that leverage optimal separating hyperplanes
(OSH) or artificial neural networks (ANN) are more accurate and
reliable.
Methods
A systematic review and meta-analysis were conducted to assess the
capability of such algorithms to predict pathological kinematic and
kinetic gait patterns as indicators of knee OA. With 153 eligible
studies, 6 studies met the inclusion criteria for a subsequent
meta-analysis, accounting for 273 healthy subjects and 313
patientswith symptomatic knee OA. The classification performance of
supervised ML classifiers (OSH- or ANN-based) used in these studies was
quantitatively assessed and compared across four following performance
metrics: classification accuracy on the test set (ACC), sensitivity
(SN), specificity (SP), and area under the receiver operating
characteristic curve (AUC).
Results
There was no statistically significant discrepancy in the ACC between
OSH- and ANN-based classifiers when dealing with kinetic and kinematic
data concurrently, as well as when considering only kinematic data.
However, there was a statistically significant difference in their SN
and SP, with the ANN-based classifiers having higher SN and SP than
OSH-based algorithms. As only one of the eligible studies reported AUC,
this metric could not be assessed statistically across studies.
Significance
This study supports the use of ANN-based algorithms for classifying knee
OA-related gait patterns as having a higher sensitivity and specificity
than OSH-based classifiers. Considering their higher reliability,
leveraging supervised ANN-based methods can aid biomechanists to
diagnose knee OA objectively.