Table 4 Performance of the classification models, broken down by each item.

From: Computer vision quantification of whole-body Parkinsonian bradykinesia using a large multi-site population

 

MDS-UPDRS classifier

Binary classification

Balanced Accuracy

Acceptable Accuracy

Accuracy

AUCROC

Finger Tapping

0.44

0.84

0.71

0.79

Hand Movements

0.43

0.86

0.74

0.81

Pronation-Supination

0.40

0.81

0.73

0.75

Toe Tapping

0.44

0.88

0.76

0.84

Leg Agility

0.52

0.91

0.80

0.86

All items

0.45

0.86

0.75

0.81

  1. For the MDS-UPDRS rating classifier; balanced accuracy (average class recall) and acceptable accuracy (proportion of predictions within ±1). For the binary classifier; accuracy and area under the receiver operator characteristic curve (AUROC). For each of these evaluation metrics, there was a small variation between items, with pronation-supination tending to perform worse, and leg agility better.
Back to article page