Subject-Specific Session-to-Session Transfer Learning Strategies for Increasing Brain-Computer Interface Performance during Upper Extremity Neurorehabilitation in Stroke

Abstract

Purpose

To assess if transfer learning strategies can improve stroke patients’ ability to control a Brain-computer interface (BCI) based on motor intention across an upper extremity neurorehabilitation intervention.

Methods

Three subject-specific session-to-session training strategies were retrospectively assessed in the present study, using information acquired during a BCI intervention in 12 stroke patients. One strategy used data from the previous therapy session (previous session), another used data from all previous sessions (accumulative) and another initially used previous session’s data and was updated with data acquired during the current session (instantaneous).

Results

Classification accuracy was significantly higher with the instantaneous strategy (median = 76.4%, IQR = [68.7%, 81.5%]) compared to the obtained with the accumulative (71.67%, [65.1%, 78.5%]) and previous session (69.2%, [59%, 77.4%]) strategies. Median classification accuracies across sessions were also higher with the instantaneous strategy in each BCI intervention session.

Conclusion

The instantaneous strategy could allow stroke patients to achieve a competitive level of BCI performance during a motor intention BCI intervention without reducing effective therapy time or requiring data from other patients.