Table 1 Evaluation metrics of all FOG classes of the top 5 models on the private (hidden) test data at the point on the precision-recall curve closest to (1,1)

From: A machine learning contest enhances automated freezing of gait detection and reveals time-of-day effects

Team

FOG class

F1 score

Accuracy

Precision

Recall

Specificity

1st place

Start Hesitation

0.459

0.996

0.346

0.681

0.997

Turn

0.814

0.920

0.821

0.807

0.951

Walk

0.079

0.682

0.041

0.915

0.678

All FOG

0.790

0.904

0.805

0.775

0.943

2nd place

Start Hesitation

0.014

0.674

0.007

0.960

0.673

Turn

0.838

0.931

0.853

0.823

0.961

Walk

0.407

0.983

0.417

0.399

0.992

All FOG

0.812

0.915

0.839

0.786

0.954

3rd place

Start Hesitation

0.439

0.997

0.445

0.434

0.999

Turn

0.843

0.933

0.855

0.832

0.961

Walk

0.053

0.476

0.027

0.972

0.469

All FOG

0.762

0.885

0.738

0.788

0.915

4th place

Start Hesitation

0.495

0.998

0.583

0.430

0.999

Turn

0.754

0.895

0.761

0.747

0.935

Walk

0.032

0.116

0.016

0.980

0.103

All FOG

0.729

0.876

0.742

0.716

0.924

5th place

Start Hesitation

0.373

0.998

0.455

0.317

0.999

Turn

0.789

0.905

0.760

0.820

0.929

Walk

0.031

0.071

0.016

1.000

0.057

All FOG

0.766

0.890

0.760

0.771

0.926

  1. Similar results were obtained for these measures when evaluating these models on the private+public test sets (see Supplementary Table 1). As mentioned above, “All FOG” does not refer to the average of the performance in each class, but rather refers to the binary case of FOG vs. non-FOG.
Back to article page