Projects from my master's research at the Rewire Lab, Department of Mechanical Engineering, University of Texas at Austin. My later work, from the PhD at Virginia Commonwealth University to wearable biosensors at NIRSense, is listed in the publications.
Building an MRI-compatible device to collect finger force data
Rewire Lab, University of Texas at Austin
To investigate how fine motor control and learning happen in the brain, I needed a keyboard and sensors that would work inside an MRI scanner (top of the first picture). The idea was to build the device from a non-metallic material. It was manufactured in Nylon 12 by laser sintering (SLS 3D printing). Each sensor is a spring-like double structure that deflects when a force is applied to the contact surface.
A second keyboard was equipped with piezoelectric force sensors (bottom of the first picture). Besides serving as a reference for its MRI-compatible counterpart, it is portable and reliable.
Nylon 12 was selected for mechanical properties that allow a very sensitive sensor. The deflection is measured in real time by an optical fiber: the distance between the head of the fiber and a mirror attached to the sensor's central part changes the light intensity, which is converted to a voltage and then to a force in newtons. All electronic parts sit outside the scanner, so the signal is not affected by the magnetic field.
This system became the experimental setup to investigate neurofeedback-based training in the MRI scanner, with the goal of improving stroke rehabilitation. My master's thesis describes the work in full.
The MRI-compatible keyboard (top) and the piezoelectric reference keyboard (bottom).
Sensor design and finite element analysis (A) and the acquisition hardware (B).Example force recording. Each color is a finger (1 for pinky to 5 for thumb). Movement time (MT) decreases with practice; overlapping fingers are visible in the second plot.
Close-up and manufacturing of the piezoelectric keyboard.Familiarization setup in the lab. Position and ergonomics match the procedure inside the scanner.
Finishing the keyboard and the command box.A key with the custom optical force sensor, inclined to match the thumb's flexion angle. Adjustable wrist padding is visible on the left.
Using machine learning classifiers to decode patterns of brain activity
fMRI multi-voxel pattern analysis
Before investigating neurofeedback for motor learning, we needed to decode the brain activity associated with executing finger sequences. The first step was to select finger sequences against specific criteria (in MATLAB):
No sequence is the same forward and backward, e.g. 1-5-3-4-2 vs. 2-4-3-5-1.
No three-number run appears in the same position, e.g. 1-5-3-4-2 vs. 2-5-3-4-1.
No three numbers match in any position, e.g. 1-5-3-4-2 vs. 1-4-3-5-2.
An even number of sequences starts with each finger, obtained by reversing some of the final sequences.
A pilot experiment on the 19 resulting sequences tested whether difficulty was consistent, as assessed by movement time and error rate.
Pilot results (n = 8; MT is movement time). Sequences in yellow were used for familiarization; those in green were used inside the scanner.Animation of the data path from a voxel to its selection in a classifier, shown as a matrix (n = classes or sequences; m = voxels).
This is a graphic representation of the classifier. SC marks every voxel chosen by selection count to be part of the classifier. Each SAT is one of the four finger sequences tested, and the color gradient is the voxel's measured activity.
This classifier reached a decoding accuracy of 58.7% when tested on samples from another scanner session, a good result for a four-class problem given that head alignment changes between sessions.
Results were weaker when testing on both hemispheres or on other regions of interest. What mattered most for decoding was the quality of the samples used to train the classifier, and further improvements were needed before applying the neurofeedback paradigm. The master's thesis has the details.
The experimental paradigm: testing whether the BOLD signal can be modulated through operant conditioning-based neurofeedback to induce motor learning.
Evaluating NIRx functional near-infrared spectroscopy to measure brain activation
fNIRS during walking and finger movements
Functional near-infrared spectroscopy (fNIRS) uses the properties of infrared light to measure local variations in blood oxygenation. It relies on an optical window, roughly 700 to 900 nm, in which skin, tissue and bone are mostly transparent to light while oxygenated and deoxygenated hemoglobin are stronger absorbers.
Like fMRI, fNIRS relies on neurovascular coupling, the blood-oxygen-level-dependent (BOLD) response, to infer changes in neuronal activity. Studies comparing fMRI and fNIRS show highly correlated results in cognitive tasks (Cui et al., 2010). I evaluated whether the NIRx system was suitable to measure movement-related brain activation.
Conditions investigated:
Symmetrical and asymmetrical walking on a treadmill.
Isolated lower-limb movements (passive vs. active hip flexion).
Finger flexion sequences.
Pros: high temporal resolution, noninvasive, portable, easy to set up. Cons: low spatial resolution, limited depth of signal, artifacts.
The walking protocol had three blocks:
Standing rest (baseline).
Asymmetrical walking, one leg at 1.3 mph and the other at 0.7 mph.
Symmetrical walking, both legs at 1 mph.
Each block lasted 30 s, and every third block returned to baseline. The order of the two walking blocks was randomized, as was which leg was slowed during asymmetrical walking. Participants watched a monitor that cued each switch a few seconds ahead (so they don't collapse!).
Timing was recorded simultaneously to synchronize brain data and behavior afterwards. Raw data were processed with the NIRx software and analyzed with SPM.
Setup for symmetrical and asymmetrical walking.
Pre-processed data roughly mapped onto the brain surface, useful to visualize the blocks along the time course.Walking experiment, SPM first-level analysis. Each map is a contrast between two conditions (e.g. symmetrical vs. rest is the increase in activation compared with rest over the whole data set). All maps except those marked with an asterisk are significant at p ≤ 0.05.Finger sequences, each compared with rest. A trend separates sequence from rest, but no differences between sequences were observed. The system's poor spatial resolution (around 1 cm) could account for this.
Electrocorticography (ECoG) recording of sequence learning
Rewire Lab and Seton Brain Institute
As part of my work on the neural correlates of motor learning, I was involved in a collaboration between the Rewire Lab and the Seton Brain Institute. Using the experimental setup I built, I collected concurrent finger force data and electrical brain activity from patients with electrodes implanted on the brain.
Pros: higher spatial resolution and signal-to-noise ratio than EEG, higher temporal resolution than fMRI. Cons: limited sampling time and field of view, invasive.
While ECoG is used primarily in epileptic patients during presurgical planning, it has also been used as a recording technique for brain-computer interface research (Leuthardt et al., 2004).
I wrote a one-page report defining the project and the protocol.
