Skip to main content

Table 1 Comparing different ML architectures and representations

From: Automatically evaluating balance using machine learning and data from a single inertial measurement unit

Representation

Test accuracy (%)

Test AUROC

Majority classifier

\(37.8\% \pm 0.0\%\)

\(0.500 \pm 0.000\)

Self-assessments

\(43.3\% \pm 0.0\%\)

\(0.665 \pm 0.000\)

Engineered features

\(\mathbf {57.2\% \pm 1.0\%}\)

\(0.768 \pm 0.007\)

Time-series

\(55.1\% \pm 2.2\%\)

\(0.801 \pm 0.004\)

Image

\(56.4\% \pm 1.3\%\)

\(\mathbf {0.806 \pm 0.002}\)

  1. All results show a clear improvement over both non ML baselines. The time-series representation as input to a 1D CNN model outperformed the engineered features in terms of AUROC, while the 2D CNN using an image representation as input resulted in the best performance. Bold signifies the best performing model for accuracy and AUROC respectively