Satyam Kumar Successfully Defended his PhD Thesis. Congratulations!!

On September 24, 2024 Satyam Kumar successfully defended his Phd thesis:

“Transferring BCI Skills : From Experts to Ecological Settings” 

Let’s Party!

We wish you all the best in your future career

Detailed Thesis Abstract:

Non-invasive brain-computer interfaces (BCI) based on electroencephalography (EEG) have proven efficient in applications such as neurorehabilitation, robotics control or immersive virtual reality. Motor imagery (MI) —mental rehearsal of a limb movement without execution— is a common EEG-BCI modality. MI elicits distinct modulations of sensorimotor rhythms (SMR) for different movements; however, online decoding is hampered by the non-stationary nature of EEG. Although complex machine learning (ML) models can alleviate this problem, subject’s learning of the BCI skill —i.e., the ability to generate distinctive SMR patterns— is also crucial to operate brain-controlled devices. Consequently, mutual learning —building ML models that promote subject’s acquisition of BCI skills— has gained increasing attention and remains an active area of interest in BCI field. Training a BCI user requires collecting individual calibration data due to high inter-subject neural variability that limits the usability of generic decoders. However, calibration is cumbersome and may produce inadequate data for building decoders, especially with naive subjects. In this thesis we explore transfer learning frameworks, across participants and across tasks, to avoid calibration pitfalls and promote subject’s learning of BCI skills.

In Part 1 of this thesis, we propose frameworks that utilize decoding models trained on data from a single expert individual that can be easily transferable to novice users through real-time incremental domain adaptation techniques. This approach allows for online BCI training without the need for individualized calibration. We evaluated our frameworks on 18 healthy na¨ıve subjects over five days, who operated a customary synchronous bar task with continuous feedback and a more challenging car racing game with asynchronous control and discrete feedback using MI BCI. We show that along with improved task-oriented BCI performance in both tasks, our frameworks promoted subjects’ ability to acquire individual BCI skills, as the initial neurophysiological control features of an expert subject evolved and became subject specific. Our presented frameworks facilitate calibration-free BCIs and have immediate implications for broader populations —such as patients with neurological pathologies— who might struggle to provide suitable initial calibration data. We then present a pilot study to test the proposed framework on two patients with stroke. We show the efficacy of using the decoder from an able-bodied expert for stroke patients to provide congruent feedback in early phases of BCI training.

One of the most promising applications of MI BCIs is in upper body rehabilitation, where patients could potentially control robotic exoskeletons using MI for closed-loop rehabilitation. However, MI decoding in the early stages often lacks robustness, requiring longitudinal training to achieve reliable MI control. Unintended and erroneous behavior from robotic exoskeletons can frustrate users and lower their interest in the rehabilitation process. To address this challenge, we characterize the neural markers associated with perceived erroneous behavior in upper body exoskeletons while naive users use MI to control the exoskeleton. Additionally, we demonstrate that these erroneous neural patterns can be successfully decoded in a subject-independent framework.

While Part 1 shows how our transfer learning frameworks can exploit the inherent consistency of neural patterns across subjects —where individual patterns are located in different parts of the input space and can be mapped to a common location to benefit from the decoder of an expert subject—, Part 2 investigates another critical aspect of the incremental online domain adaptation at the core of our transfer learning frameworks. Since our algorithms help minimize shifts in feature distributions, they could also cope with non-stationraities across tasks. Part 2 demonstrates the effectiveness of our adaptive frameworks in real-life ecological settings involving dual tasking. In these settings, the challenge of non-stationarity in neural patterns is further exacerbated by the presence of background neural activity from concurrent tasks. Similar to the consistency observed in MI patterns across subjects that allows inter-subject transfer learning, we hypothesize that the neural patterns of MI remain consistent when performed alone versus when simultaneously executed with other tasks.

Walking, a fundamental activity of daily living, is the first domain to showcase the robustness and transferability of the adaptive decoding frameworks. We show that a MI decoder trained on data collected while subjects are seated imagining left and right hand movements can provide reliable online feedback while those subjects perform hand MI and simultaneously walk on a treadmill. Additionally, to explore the integration of BCI systems with residual or functional limbs, we show that these adaptive frameworks facilitate the execution of bimanual tasks requiring the simultaneous operation of a brain-controlled hand exoskeleton and overt motor activity from the opposite functional hand. Our findings highlight the necessity of longitudinal training for reliable BCI performance. We observed that while BCI control initially decreases when transitioning to bimanual tasks as compared to just the BCI tasks, performance improves with multi-day training, and that the improved performance are grounded in the users’ ability to generate increasingly discriminant patterns over time.