Initial motor skill performance predicts future performance, but not learning

Abstract

People show vast variability in skill performance and learning. What determines a person's individual performance and learning ability? In this study we explored the possibility to predict participants' future performance and learning, based on their behavior during initial skill acquisition. We recruited a large online multi-session sample of participants performing a sequential tapping skill learning task. We used machine learning to predict future performance and learning from raw data acquired during initial skill acquisition, and from engineered features calculated from the raw data. Strong correlations were observed between initial and final performance, and individual learning was not predicted. While canonical experimental tasks developed and selected to detect average effects may constrain insights regarding individual variability, development of novel tasks may shed light on the underlying mechanism of individual skill learning, relevant for real-life scenarios.

Figure 1

Task performance within and between sessions. (a) Experimental design. (b) Learning curves across all five sessions (session 1—blue, session 2—yellow, session 3—green, Retention session—orange, Long Term Retention session—pink), the shaded area represents the 95% confidence interval. (c) Offline gains between consecutive sessions. Data points in the violin plots represent offline gains for each participant. The white dot represents the median.

Figure 2

Model performance with engineered features. (a) Maximum mean cross-validation R² scores (blue) and the corresponding hold-out R² scores (orange) for each learning interval (x axis). (b) Maximum mean cross-validation R² scores (blue) and the corresponding hold out R² (orange) for the two retention intervals (x axis).

Figure 3

Performance was consistent across sessions but did not predict learning. (a) Performance in all sessions explains a large portion of the variability in future performance (R² range = [0.25, 0.91]. (b) Performance hardly explains the variability in learning (R² range = [0, 0.05]. (c) Performance throughout the experiment separated according to the performance quantile in the Retention session (colors), showing that participants' relative performance rank remains stable across sessions. Shaded areas represent the 95% confidence interval. Statistical significance is marked with * for p < 0.05 and with ** for p < 0.001.