📝 Research

EEG/ERP Data Preprocessing: Systematically processed resting-state EEG and N-back task ERP signals using EEGLAB, implementing a full pipeline of filtering, artifact rejection, and normalization to maximize signal integrity and ensure high-quality data for analysis.

Frequency-Domain Biomarker Identification: Employed Fourier Transform to extract power and energy spectrum features, converting complex time-series data into a clear frequency-domain representation to uncover potential biomarkers of schizophrenia.

Interpretable Machine Learning for Diagnosis: Developed diagnostic models using interpretable algorithms (SVM, GBDT, KNN, Random Forest). Leveraged the Fisher-score algorithm for feature selection to pinpoint the most discriminative neural patterns across different frequency bands and brain regions.

Key Findings: Demonstrated that GBDT model achieved superior performance in classifying resting-state EEG, while SVM was optimal for task-based ERP data. This highlights a state-dependent optimal model choice. Identified the occipital lobe during visual processing tasks as a critical region containing highly discriminative features, suggesting its key role in the pathophysiology of schizophrenia.

Cross-Cohort & Cross-Age Comparative Framework: Initiated a novel comparative study between a privately collected adult schizophrenia EEG dataset (Shanghai) and a public adolescent dataset (Moscow), establishing a framework to test the generalizability of neural biomarkers across different age groups and populations.

Functional Brain Network Construction & Analysis: Constructed functional brain networks using Power Spectral Density (PSD) as node features and Phase Lag Index (PLI) / Phase Locking Value (PLV) as edge weights. This approach enabled a direct comparison of functional connectivity patterns between the two distinct patient cohorts.

Graph-Based Deep Learning for Neuropathology: Applied Graph Convolutional Network to model the brain's functional connectome, successfully identifying complex, non-linear neuropathological patterns that distinguish the adult and adolescent schizophrenia cohorts.

Developing a Superior Hybrid Diagnostic Model: Designed and benchmarked multiple deep learning architectures for diagnostic classification. The proposed LSTM-Transformer model demonstrated superior performance and robustness, outperforming other hybrid models across both datasets, providing a powerful solution for cross-cohort diagnosis.

A Holistic, LLM-Enhanced Diagnostic Framework: Drawing on my industry experience with Large Language Models (LLMs), I engineered a novel platform that integrates three critical data sources: objective EEG-based diagnostic predictions, standardized clinical assessment scales, and a dynamic, interactive patient consultation module.

Rapid Prototyping and Research-to-Practice Translation: I demonstrated strong self-learning and execution capabilities by mastering the Streamlit framework from the ground up to build and deploy a functional prototype in just two weeks. This platform is specifically designed to create a seamless workflow between computational research and hands-on clinical assessment.

Domain-Specific Language Model Customization: To power the platform’s consultation module, I built a specialized, lightweight language model. By leveraging Retrieval-Augmented Generation (RAG), the model was fine-tuned on a curated corpus of medical literature and clinical reports to provide context-aware and evidence-based responses for clinical inquiries.

Domain-Specific Text Processing Pipeline: Engineered a specialized text processing pipeline to handle the unique challenges of medical literature. By training a word segmentation model on manually annotated samples, the system achieves high accuracy in identifying professional medical terms within a large-scale corpus of stroke research abstracts.

Head-to-Head Comparison of Word Embedding Models: Conducted a direct comparative analysis of two seminal word embedding algorithms, GloVe and Word2Vec. Both models were trained on the processed corpus to create distinct vector-space representations of the stroke research domain, allowing for a rigorous evaluation of their ability to capture semantic relationships.

Quantitative and Qualitative Performance Evaluation:Employed unsupervised learning (cluster analysis) as a downstream task to provide an evidence-based evaluation of each vectorization model. By comparing the quality of the resulting thematic clusters, this work offers a clear conclusion on which algorithm is superior for uncovering meaningful, latent structures within specialized scientific literature.

Systematic Mapping of the Global Research Landscape: Conducted a large-scale bibliometric analysis to identify key research trends, geographical patterns, and influential institutions and authors in the field of DCD intervention. The study highlights Canada, particularly McMaster University, as a central hub for DCD research.

Evidence-based Evaluation of Intervention Efficacy: Performed a meta-analysis to quantitatively assess the effectiveness of various motor interventions using the standardized Movement Assessment Battery for Children-2 (MABC-2).

Identification of Optimal Intervention Strategies: The findings demonstrate that different interventions yield distinct benefits. Notably, multi-component exercises are most effective for improving balance, while task-oriented and skill-based activities show significant advantages for enhancing manual dexterity, aiming, and catching.

Guidance for Future Clinical Practice and Research: By integrating bibliometric trends with quantitative efficacy data, this research identifies critical gaps in the literature and proposes a clear, data-driven framework for designing and implementing targeted DCD management strategies.

FPGA-Based Real-Time Algorithm Implementation: A fatigue detection algorithm was designed and implemented on an FPGA. This on-chip processing enables real-time signal analysis at the edge for a wearable predictive warning system.

Non-Linear Feature Engineering: A feature extraction pipeline was developed to quantify fatigue from raw EEG signals. The process involves computing non-linear dynamic metrics, including Sample Entropy and Approximate Entropy, to characterize changes in brain activity associated with physiological fatigue.

Dual-Approach Deep Learning for Classification: Two deep learning models were architected and benchmarked for fatigue classification. A CNN-LSTM fusion model was applied to capture spatio-temporal dependencies in EEG signals, while GNN was used to model the brain's functional connectivity.