Brain-based digital twins for skilled interceptive movements: computational frameworks for neural-kinematic integration

Abstract

Human motor imagery provides a window on how the brain plans and refines movement, but its integration into digital-twin technology for real-time skill improvement remains limited. This thesis addresses that gap by formalising a brain-based digital twin (BB-DT) for cricket batting motor imagery that fuses electro encephalography with synchronised kinematic data to model, predict and provide feedback on skilled interceptive actions. The thesis develops a four-layer neuro-kinematic architecture spanning signal processing, synthetic-data augmentation, personalised neural mass modelling and closed-loop intervention. It specifies validation criteria - representational fidelity, identifiability and tractability - for judging when a brain-based twin is theoretically sound, and treats synthetic EEG as a theory-guided instrument for hypothesis testing rather than a simple data surrogate. Methodologically, the research combines consumer-grade 14-channel EEG, expert-validated cricket-stroke videos, kinematic capture, and a matched-filter signal-processing pipeline. A conditional GAN supplies 85.1% synthetic augmentation to mitigate empirical sparsity, while an XGBoost classifier delivers real-time mental-state discrimination within an 80 ms feedback window. The BB-DT is demonstrated in a proof-of-concept coaching interface that provides neural efficiency feedback during batting imagery. The evaluation combines the cricket-specific dataset with external EEG datasets so that the computational claims are tested against both domain-specific and cross-dataset evidence. Across the thesis, the framework is assessed through controlled participant recordings, synthetic-data validation, latency profiling, classification experiments and prototype implementation evidence. This combination addresses central constraints in sport EEG: limited labelled data, noisy portable recordings, synchronisation between neural and movement streams, and the need for feedback fast enough to support training-oriented use. Findings show that domain-constrained synthetic augmentation maintains classification accuracy within 2% of all-real models and supports stable personalisation across sessions. The work also shows how a computational twin can expose interpretable links between alpha-band desynchronisation, predicted bat trajectory and subsequent kinematic refinement, extending prediction-based accounts of motor control. The thesis positions these results as proof-of-concept evidence rather than field-ready performance enhancement, with the principal validation focused on feasibility, latency, data governance and interpretable model behaviour under controlled laboratory conditions. The principal contribution is a unified computational framework for skilled interceptive movement that integrates streaming neurophysiology and kinematics for real-time modelling, classification and feedback within practical system constraints. A secondary contribution is a data-engineering and evaluation strategy for sparse sport-specific EEG settings, combining governed multimodal datasets, synthetic augmentation and reproducible cross-dataset validation. Together, these contributions provide design principles that generalise to rehabilitation, education, training technology and creative computing contexts where neural, behavioural and contextual data must be integrated responsibly.

The authors' abstract, as published at the source. ResearchSPAce (Bath Spa University), 2028 · DOI ↗