Fig. 3: Results of the best-performing classification experiment presented in a color-coded matrix depicting the performance score and the employed time series or combination of such.
From: Leveraging machine learning for digital gait analysis in ataxia using sensor-free motion capture
Results of tsfresh+XGBoost classification model where each row-column combination in the upper right triangle depicts the best reported macro-averaged F1-score (in %) for the respective binary classification. The lower left triangle depicts for which time series or combination of time series this performance was reported. Classes labeled 0–4 refer to patient data only, while the class labeled HC includes all healthy control subjects. For instance, the best tsfresh+XGBoost model fed with time series features derived from the x-positions (X-pos) and angles at the hips (Lower) and trained to classify between the SARA gait scores 0 and 1 within the ataxia group was able to score a macro-averaged F1-score of 61.73%. X = X-pos (time series of raw x-positions of each marker separately), D = Dist (time series of distances between two markers), U = Upper (time series of angles of the upper body part, i.g. shoulders), and L = Lower (time series of angles of the lower body part, i.g., hips). HC = Healthy Control.
