End-to-end bimanual motor skill assessment from raw neuroimaging data: deep learning and interpretable cross-procedural generalization
Loading...
Authors
Subedi, Aseem, Pratap
Issue Date
2025-12
Type
Electronic thesis
Thesis
Thesis
Language
en_US
Keywords
Mechanical engineering
Alternative Title
Abstract
Objective neuroimaging-based assessment of bimanual motor skills faces three critical methodological challenges: (1) dependence on extensive preprocessing and manual feature engineering, which limits real-time applicability and risks obscuring relevant neural dynamics; (2) the lack of inherent interpretability in complex data-driven models, which restricts actionable insight and clinical trust; and (3) limited generalizability, as existing models often fail under out-of-distribution conditions and require substantial retraining for new users, tasks, or procedural domains. This dissertation addresses these gaps through two independent aims.Aim 1 develops and validates an end-to-end deep learning framework for real-time, objective assessment of bimanual motor skills across three distinct, clinically relevant tasks—Fundamentals of Laparoscopic Surgery (FLS) Suturing, FLS Pattern Cutting, and endotracheal intubation (ETI)—using raw functional near-infrared spectroscopy (fNIRS) signals. A one-dimensional convolutional neural network (CNN) achieves high classification accuracy (up to 98.6%) and a robust Matthews correlation coefficient (up to 0.917) across all tasks and relevant brain regions, significantly outperforming preprocessing-dependent methods (p<0.05). Subject-level analyses reveal temporal dynamics of cortical activation linked to skill consolidation, with the highest predictability observed after the fifth day of training.
Aim 2 establishes an interpretable, transformer-based foundation model for cross-procedural generalization, leveraging self-supervised pretraining on two of the aforementioned tasks and then adapting to a fourth, previously unseen emergency procedure: cricothyrotomy. The model achieves binary classification AUCs of 0.83 and 0.89 on two cricothyrotomy studies using fewer than 30 labeled samples, enabled by a lightweight (< 2,000 parameter) adapter module. Interpretability is embedded through channel- and temporal-attention mechanisms, highlighting task-relevant prefrontal sub-networks—particularly in the dorsolateral prefrontal cortex (DLPFC) and frontopolar regions—validated by ablation analyses and alignment with established neurophysiological patterns.
Together, these innovations establish a scalable and interpretable foundation for real-time, objective neuroimaging-based skill assessment, thereby advancing the state-of-the-art in performance monitoring and training across medical and human performance domains.
Description
December2025
School of Engineering
School of Engineering
Full Citation
Publisher
Rensselaer Polytechnic Institute, Troy, NY