Table 3 Cross-validated performance for three motor performance-related classification problems

From: Integrating digital gait data with metabolomics and clinical data to predict outcomes in Parkinson’s disease

Models

Linear SVM

RBF SVM

RF

Stochastic GBM

DEEP

XGB

Clinical confounders (median)

10-fold CV

AUC

median

(mad)

FoG

0.694

(0.04)

0.792

(0.19)

0.889

(0.11)

a0.917

(0.12)

0.833

(0.12)

0.903

(0.14)

0.609

(0.11)

Gait disorder occurrence

a0.742

(0.14)

0.562

(0.20)

0.589

(0.11)

0.569

(0.18)

0.577

(0.17)

0.652

(0.22)

0.554

(0.13)

PDQ-39 mobility sub-score

0.646

(0.12)

0.781

(0.09)

0.759

(0.19)

0.795

(0.16)

a0.799

(0.14)

0.764

(0.12)

0.535

(0.09)

  1. Cross-validated performance for three classification problems using different machine learning methods and time series features computed from the raw gait signal data as input: (1) Freezing of Gait (FoG) score binary classification into scores above/below the median; (2) gait disorder occurrence prediction (yes/no); (3) predicting whether the PDQ-39 mobility sub-score of PD patients is above the median in the cohort (yes/no). “Clinical confounders” refers to a model that was solely trained with age and sex as predictors and serves as a comparator.
  2. AUC area under the Receiver Operating Characteristic Curve, median and median absolute deviation (mad) across 10 cross-validation (CV) cycles, SVM support vector machine, RBF radial basis function, DEEP deep boosting, XGB extreme gradient boosting, RF random forest, GBM gradient boosting machines.
  3. aThe highest median AUC for each row.
Back to article page