Fig 1.
A schematic comparison between (a) a hypothesis-testing experimental paradigm and (b) a data-driven machine learning analysis.
Fig 2.
Schematic image of ML-based pitch-type prediction using joint angles at time t as features.
Fig 3.
Schematic illustration of the two analyses conducted in this study.
(a) Analysis 1: Sliding time-window analysis for identifying spatiotemporally informative cues for the ML model. (b) Analysis 2: Set-size analysis for evaluating how the accumulation of information across trials influences the prediction accuracy of the ML model.
Table 1.
Summary of the fastballs and breaking-balls information for each pitcher.
Fig 4.
Illustrations at 20% intervals of the normalized motion, using Sub1 as an example.
Fig 5.
Mean prediction accuracy and standard deviation for each pitcher at each time point using full-joint information.
Panels (a)–(h) show the individual plots for subjects 1–8.
Fig 6.
Overall mean prediction accuracy and standard deviation across the eight pitchers using full-joint information.
Fig 7.
Mean prediction accuracy for each of the six body regions, with results from all eight pitchers overlaid.
Fig 8.
Mean prediction accuracy and standard deviation across eight pitchers for each of the six body regions.
Solid lines indicate significant intervals (p < .05).
Table 2.
Relationship between each parameter and the prediction accuracy of each pitcher.
Fig 9.
(a) Comparison of prediction accuracy among logistic regression, GRU, and LSTM models.
(b) Comparison of prediction accuracy of models trained using (i) joint angle information, (ii) joint angles and angular velocity, and (iii) joint angles, angular velocity, and angular acceleration.
Fig 10.
Relationship between dataset size and prediction accuracy at each time point for each pitcher.
Panels (a)–(h) show individual plots for subjects 1–8.
Fig 11.
Improvement in mean prediction accuracy with increasing training data.
Numbers on the dataset size axis represent the number of trials per pitch type (thus, a size of 5 corresponds to 10 training trials in total). Solid lines indicate intervals where accuracy improvements were statistically significant compared with the chance level (0.0).