Pain neuroscience and machine learning

My PhD combines graded ulnar-nerve stimulation in anaesthetised pigs with simultaneous intracranial recordings from S1, PFC, and an ACC-targeted depth array. We developed an approach to ACC access and electrode targeting using surgical landmarks and real-time electrophysiological guidance. I performed craniotomies, durotomies, and surface- and depth-electrode placement to acquire these recordings. I develop the supporting data and quality-control workflow, analyse cortical responses in time and frequency, and build interpretable deep-learning models with animal-level evaluation.

Our Tucker-Davis Technologies (TDT) recordings did not fit neatly into the Python tools and data-sharing workflow I wanted to use. I started building a converter around the Brain Imaging Data Structure (BIDS), adapting its organisation to our animal recordings.

The pipeline reads TDT data into MNE-Python and keeps S1/PFC and ACC recordings in separate files to accommodate their different sampling rates. It currently saves FIF files with recording metadata and stimulation events in a BIDS-inspired layout, keeping recordings separate from processed outputs.

This is still in development. My aim is to make it reusable beyond our dataset, helping other TDT users prepare recordings for analysis and open-data sharing.

At the Grill Lab at Duke, I investigated how spinal cord stimulation waveforms change neural-fibre activation. Using Python and PyFibers, I simulated myelinated axons with an interpolated McIntyre–Richardson–Grill (MRG) model, representing nodal ion-channel dynamics and current flow through the myelinated fibre.

I compared tonic, FAST, and burst stimulation across fibre diameters, pulse widths, and active or passive charge balancing. I used bisection searches to estimate activation thresholds, then swept stimulation amplitude to examine how reliably each pulse produced an action potential. Spatiotemporal voltage maps let me inspect where spikes began, how they propagated, and when burst responses became irregular.

The simulations showed lower activation thresholds but less consistent firing with burst stimulation. This work became an IEEE EMBC 2025 paper.

The model explores possible mechanisms rather than predicting clinical outcomes.

I wanted to check whether a card in a shop was worth buying without manually searching several marketplaces. I used a pretrained DINOv3 vision transformer to index reference-image embeddings and retrieve matches by cosine similarity, with OCR to help resolve card identity. I also used 3D UMAP to inspect the embedding space.

After the user confirms a match, the app brings together available marketplace API prices and compares ungraded resale and graded-card scenarios against the sticker price, selling fees, grading costs, and configurable margin thresholds. Missing prices can be entered manually. It is a hobby project for comparing purchase options, not predicting resale returns.