2025 Summer uPNC
Project Description:
Spinal Muscular Atrophy (SMA) is a genetic neurodegenerative disorder caused by reduced levels of the survival motor neuron (SMN) protein, leading to motor neuron dysfunction, muscle weakness, and fatigue. Although new SMN- enhancing therapies have improved survival and slowed disease progression, fatigue remains one of the most reported and disabling symptoms in SMA. Current clinical assessments of fatigue rely on submaximal functional tasks, which often conflate fatigue with muscle weakness, limiting their ability to isolate neural contributions to fatigability.
This summer, Zarif Cabrera developed a standardized, robot-assisted dynamic task to quantify upper limb fatigability in non-ambulatory individuals with SMA. The protocol required repeated maximal voluntary elbow flexion and extension at a constant angular velocity using a robotic dynamometer. To ensure consistency and engagement across trials, Zarif created a custom graphical user interface (GUI) that provided real-time visual feedback, including live torque visualization and target levels for each contraction. The GUI operated by reading analog output from the HUMAC NORM Dynamometer through a National Instruments Data Acquisition Device. Baseline torque readings for each phase were used to normalize offsets due to the weight of the participant’s arm or noise. Target contractions were set using best-fit curves for flexion and extension MVC, and all data were logged to CSV files for post-processing.
Preliminary feasibility testing in one non-ambulatory SMA participant demonstrated successful implementation of the protocol, with measurable declines in elbow torque following the fatiguing sequence (26% in flexion and 20% in extension). The task was well tolerated, and continuous torque data were collected without adverse events. These findings indicate that this dynamic approach provides a sensitive, standardized measure of performance fatigability during functionally relevant movements. Future work will expand testing to healthy controls to compare fatigue profiles and determine whether fatigue in SMA reflects an independent physiological phenomenon rather than secondary effect of muscle weakness. Ultimately, this method lays the groundwork for integrating neurophysiological measures such as transcranial magnetic stimulation (TMS) and high-density EMG to investigate mechanisms of fatigue and evaluate the efficacy of emerging interventions, including spinal cord stimulation.
Project Description: Adaptation of spinal reflexes during split-belt walking
Walking usually feels automatic, yet people who have had a stroke must relearn this skill. Christian’s summer project in Dr. Gelsy Torres-Oviedo’s Sensorimotor Learning Lab examined how the spinal cord helps automate a brand-new gait when each leg moves at a different speed on a split-belt treadmill. Using gentle electrical stimulation of the tibial nerve while participants walked, we tracked the size of the soleus H-reflex, which shows us a window into spinal excitability, and related it to changes in step-length symmetry. Reflexes were noticeably smaller the moment the u familiar task began, then rose toward their original size once the pattern felt natural again, paralleling improvements in symmetry. These trends confirm that spinal circuits recalibrate as practice drives gait automaticity.
The specific aspects of the research conducted by Christian are as follows:
- Motion capture and real-time metrics – I placed reflective markers on each participant and operated a multi-camera Vicon Nexus system to record 3-D kinematics. I computed step-length asymmetry so we could quantify learning and adaptation
during the experiment. - Electrophysiology – I applied Delsys surface EMG electrodes, positioned stimulating electrodes, and delivered carefully timed pulses to evoke high-quality H-reflexes in the soleus muscle.
- Automated stimulation calibration – After each calibration trial, I generated recruitment curves for the M-wave and H-wave, wrote MATLAB scripts that pinpoint the optimal stimulation amplitude, and integrated this routine into the collection pipeline, improving accuracy and reducing setup time.
- Data processing and visualization – I converted Vicon C3D files to MATLAB format with custom batch scripts, debugged the entire codebase, and produced the figures featured on my poster.
- Full-protocol data collection – I recruited volunteers, led each step of the roughly five-hour protocol, synchronized treadmill, EMG, and kinematic systems, and recorded auxiliary prefrontal-cortex signals with an fNIRS head-band for future analysis.
By showing that spinal reflexes are suppressed during the first encounters with a novel gait and rebound once the pattern is mastered, this work highlights the spine’s role in making locomotion automatic. These insights can guide therapists in designing spinal-targeted interventions that help stroke survivors regain smooth, symmetric, energy-efficient walking.
Project Description:
Human vision shows an “eccentricity bias" where central (foveal) vision supports fine detail such as faces and text, while peripheral vision captures larger objects and broad scene layout. This phenomenon is observed neurally in fMRI, yet it remains debated whether the bias arises primarily from intrinsic sampling constraints or is learned externally from natural viewing (eye movements, object size statistics, task demands).
Over the summer, Dylan implemented a dual stream DCNN to test whether architectural constraints plus viewing statistics are sufficient to reproduce the bias. Both streams were pre trained with a self supervised contrastive objective (SimCLR) and then fine tuned on a semantic classification task. The model comprises a high-acuity foveal stream and a low-acuity peripheral stream whose features are fused downstream alongside single stream heads. We trained on egocentric video from the Visual Experience Database (VEDB) under two conditions: Gaze-TTM and Center-Blur. In Gaze-TTM, a foveal crop was applied per frame based on the participant’s gaze, and the surrounding periphery was transformed with Freeman and Simoncelli’s Texture Tiling Model (TTM), a more biologically grounded proxy for peripheral vision than simple blur. Center-Blur served as a control without gaze-contingent crops and with a Gaussian-blurred periphery.
On a small sample (n=10 videos, ~5,000 frames), the foveal stream performed better during SimCLR pre-training metrics, whereas the peripheral stream performed better on the downstream classification task; fusion generally outperformed single streams. These patterns held for both conditions. With a larger SimCLR sample (n=266 videos, ~103,600 frames), the foveal stream showed a preference for small, detail heavy categories when evaluated on a representative set of fine vs. broad tasks.
These preliminary results suggest model level alignment with the expected performance pattern of eccentricity bias; next steps include scaling peripheral and fused evaluations to the larger sample, testing additional model configurations and frame conditions (e.g., baselines and controls to rule out trivial explanations), and, most importantly, assessing brain like similarity via RSA/encoding against fMRI data from the Natural Scenes Dataset (NSD).
Project Description:
Under the mentorship of Dr. Lee Fisher and Dr. Robert Gaunt, Andrew Dirkse studied blockage of action potentials in peripheral nerves using an ultra-low frequency (ULF) waveform current. Spinal neuromodulation using ULF current has previously been found to block axonal action potential propagation and reduce chronic pain. Andrew designed a MATLAB + Excel pipeline to derive insights into the effects of ULF current on the nerve from physiological data. Andrew also developed a computational model in the NEURON simulation environment through a Python (+ Numpy, Scipy, Matplotlib) interface. The model was based on a previously published (McIntyre-Richardson-Grill) model of a myelinated axon. Conduction block was observed during both phases of the waveform at higher amplitudes of current but incompletely (only during anodal phases) at lower amplitudes. The mechanism proposed to block propagation of action potentials was alternating opening/closing of sodium ion channel activation/inactivation gates. Overall, this study suggests that ULF can successfully block action potential propagation in peripheral nerves and highlights several avenues for future research.
Project Description:
Several studies have shown social interactions to be rewarding via striatal activations through dopaminergic fibers. However, it remains unclear how social interaction and dopamine release plays out in Shank3 KO pups. In mice, loss of function of Shank3 gene serves as a model for autism and the social deficits associated with it. Knowing that the pup dopamine system is functional in P15 pup slices and that the dopaminergic neurons of the Ventral Tegmental Area (VTA) system are also involved in social interaction, we recorded dopamine levels in ventral portion of the striatum in both WT and Shank3 KO pre-weaning mice interacting freely with their mothers to investigate how neurodevelopmental disorders might influence the rewarding nature of dopamine under ethologically rewarding conditions. We used DeepLabCut software to develop a positional tracking model that was able to track dam-pup interactions to further develop our analyses. Interestingly, we found that both WT and Shank3 KO mice exhibit pronounced dopamine transients in the absence of social interaction before the dam was introduced; additionally, maternal interactions seem to drive the dopamine transients in both WT and Shank3 KO mice after dam introduction. Notably, Shank3 KO pups display significantly higher amplitude and broader dopamine peaks during maternal interactions relative to WT mice, a difference specific to maternal bout-associated peaks.
Project Description:
Curvilinear and rectilinear features are both shown to be important for visual perception. Curvilinear features, unlike rectilinear features, are sufficient for object recognition as well as discrimination between inanimate and animate objects. Under the guidance of Dr. Maggie Henderson, Will explored how these features vary with physical size, how these features are used by classifiers, and where in the brain these features are used.
We began by extracting curvilinear feature values and rectilinear feature values using oriented Gabor wavelets, some straight for rectilinear features and some curvy for curvilinear features. We did this specifically within object masks that we got from Microsoft’s COCO dataset. These object masks were also labeled by object type which allowed us to sort the objects by physical size. We also binned the objects by their visual size to avoid that as a confound, then fed the feature values as well as the physical size to both the dimensionality reduction model t-SNE and a random forest classifier. Both models showed some ability to discriminate physical size based off the curvilinear and rectilinear features. The random forest model showed a heavy reliance on the curvilinear features for its classification.
Previous fMRI studies have shown regions of the brain that have preferential bias to curvilinearity, and we were curious about this effect in the context of the same natural scenes stimuli we were feeding our classifier model. We used the Natural Scenes Dataset: a large fMRI dataset that showed participants images of natural scenes from Microsoft’s COCO dataset. Using this dataset, we built a population receptive field (pRF) encoding model, which predicts single-voxel responses based on estimated neural populations tuned to specific visual field locations. In one of the models we only included pRF regions that overlapped with the object masks and in another we used the normal pRF regions. We found, using a variance partitioning analysis, that in both models curvilinear features uniquely explain more variance than rectilinear features. This shows that like the classifier, our brains are heavily reliant on these curvature features for visual perception.
Project Description:
Under the mentorship of Dr. Bard Ermentrout, I designed and implemented a color-extended Wilson–Cowan neural field model to explore how flicker-induced visual hallucinations emerge from dynamic instabilities in cortical E/I circuits. Building directly on the spatiotemporal framework of Rule et al. (2021) and incorporating ideas from Faugueras et al. (2022), I created a ring-based architecture in which multiple color channels (e.g., red–green, later extended to eight hues) each contain excitatory and inhibitory populations with spatially localized Gaussian excitation and color opponent inhibition shifted by π in color space. This allowed me to test how cross-color coupling and flicker parameters shape the emergence of entoptic-like patterns. In the two-color, two-location model, for example, Green Left (ue1) and Red Left (ue3) populations developed strong anti phase oscillations after ~1000 ms, indicating a color pattern rather than a purely spatial one. In the eight-color ring and 2D extensions, flicker periods of 45–65 ms and 105–125 ms produced the strongest alterations across channels, whereas other periods yielded monochromatic or weakly modulated activity.
Going forward, I plan to systematically sweep inhibitory coupling (c_i), excitatory spread (σ_e), and amplitude/per period combinations, quantify emergent limit cycles using measures such as phase coherence and color entropy, and extend the model to higher-dimensional and biologically calibrated regimes. These steps will allow for formal bifurcation analyses and comparison of the model’s predictions to psychophysical data on flicker-induced phosphenes.
Project Description:
Transcranial Electrical Stimulation (TES) is a non-invasive neuromodulation technique that can be used to treat chronic conditions such as fibromyalgia or depression. One clinical procedure during patient treatment in this lab is confirming that TES on M1 is properly administered. This is done by reaching motor threshold, which is typically verified by detecting Motor Evoked Potentials (MEPs) in EMG recordings. However, MEPs can be weak or noisy, making visual confirmation unreliable. This project aims to develop a robust classifier to automatically detect MEP responses and determine whether motor threshold was reached, even in high noise-to-signal ratio conditions.
EMG data was collected from four healthy participants during TES, from flexor and extensor muscles in the dominant forearm. A total of 407 pre-labeled recordings (284 suprathreshold, 123 subthreshold) were used. After preprocessing with bandpass (25–400 Hz) and notch (60 Hz) filtering, 100 ms time windows around the TES delivery were analyzed. Features were extracted to capture wave characteristics such as amplitude, peak shape, frequency content, and signal variability. These features were used to train supervised machine learning models, including optimized neural networks.
Cross-validation and testing results showed area under the ROC curve (by AUC) exceeding 90% across all optimized models. This was also true for F1 scores. Neural networks demonstrated the most consistent performance. Shapley importance revealed that out of the 10 features extracted, about 4 specific features drove the majority of classification accuracy. While binary threshold detection scores high performance, the classification of which muscle groups the MEP was recorded from, was less reliable (~85% accuracy), suggesting a need for specialized features and more training data.
Future work will focus on refining classification near threshold boundaries and applying this system to EMG data from fibromyalgia patients unde going TES.
Nicholas Liotta
Undergraduate Institution: University of Buffalo
Mentor: Jonathan Tsay
University and Department: 鶹
Project Description:
Project Description:
Locomotor adaptation allows individuals to adjust both movement and perception in response to novel walking conditions. One method of studying this is the split-belt treadmill, where each leg moves at a different speed. As people adapt, they not only change how they walk but also shift their perception of symmetry, reporting equal leg speeds even when an actual difference remains.
This study examines when that shift in perceived symmetry occurs using a Hidden Markov Model (HMM). Participants walked on a split belt treadmill while periodically reporting which leg felt slower. These responses were used to infer a latent perceptual state such as “baseline,” “adaptation,” or “post-adaptation.” Unlike traditional analyses that average across trials or conditions, the HMM captures trial-by-trial transitions in perception, allowing us to track when someone perceives equal leg speeds, not just if that perception changes.
Our results show that perceptual adaptation occurs gradually, with HMMs reliably detecting the emergence of a 60% change in perceived symmetry after the baseline measurements and before steady behavioral patterns appear. This approach refines our understanding of sensory adaptation during walking. It also highlights the value of modeling temporal dynamics in perceptual change, such as identifying when stroke patients begin to relearn normal walking patterns, guiding more personalized and successful rehabilitation.
Project Description:
Inhibitory neurons known as parvalbumin interneurons (PVIs) play a crucial role in cortical processing, and intriguingly, they fire differently depending on brain region: some fire with a delay (dFS), while others fire continuously (cFS). To investigate the mechanistic basis of the observed firing differences, a well-established mathematical model of fast-spiking interneurons was modified to incorporate an A-type potassium current, consistent with experimentally characterized Kv1.1 channel activity.
Claire's summer work focused on systematically analyzing the relationship between and among key parameters (e.g. leak conductance, A type potassium current inactivation, etc.) and their effects on firing behavior. This intensive modeling effort allowed for the reproduction of the delayed and continuous spiking patterns observed experimentally.
With these individual neuron models, Claire, Dr. Ermentrout (PI), Dr. Guillermo Gonzalez-Burgos will further collaborate to construct a neuron network to test the effects of the naturally occurring ratios of the cFS and dFS on gamma oscillations.
Project Description:
Serotonin is a well-known neuromodulator often recognized for its role in mood; however, it also plays a crucial role in reward-related behaviors such as anticipation, valuation, and consumption (Spring & Nautiyal, 2024). Cavaccini et al. (2018) further demonstrated that activation of the 5-HT4 serotonin receptor modulates synaptic plasticity at thalamostriatal synapses of direct pathway spiny projection neurons (dSPNs) in transgenic DREADD (Designer Receptors Exclusively Activated by Designer Drugs) mice. Both studies independently show that serotonin exerts input-specific modulation of dSPNs, acting through striatal circuits independently of dopamine. This project builds on those findings by using a modified version of the CBGTPy (Cortico-Basal Ganglia-Thalamic Python) model created in Clapp et al. (2025), with an added serotonin module targeting the thalamic-to-dSPN pathway. The primary aim is to simulate synaptic plasticity outcomes consistent with Cavaccini et al. (2018), supported by the behavioral relevance of serotonin dynamics described in Spring & Nautiyal (2024). Experimental findings were successfully recreated, indicating the potential for a computational model of serotonergic modulation on the thalamostratal pathway. Future research can use the model created in this study to further analyze how serotonin affects these pathways, as well as interactions between serotonin and dopamine.
Project Description:
The motor system, one of the core components of the integrated nervous system, is well-studied but not fully characterized. Neural systems coding for vision have been shown to have preferences for orientation. Do we have reason to believe the same is true for motor neural systems? Intuitively, it is reasonable that different external environments, in particular with different common orientations, might lead to encoding of different neural mechanisms for motor movements. If the different types of information that modulate motor learning could be characterized, fundamental advancements in human and robotic movement sciences as well as rehabilitation would be accelerated.
To serve exactly this question, this project examines both slow and fast implicit learning processes in the context of a reaching experiment with two distinct target orientation groups: cardinal and diagonal. Following data collection which itself was implemented and carried out during the summer, a comprehensive preprocessing pipeline was developed using Python to characterize highly noisy online motor data. A linear Gaussian state-space model, essentially a hidden Markov model with latent variables continuous in ℝ2 space, was further designed, implemented, and fit to experimental data to predict learning parameters. Only the error sensitivity, akin to a relative learning rate, differed significantly between groups, supporting experimental results. In both model predictions and experimental data, a difference in magnitude of a slow long-term implicit process is found responsible for a difference in motor learning across distinct orientation categories. The direction of a target in 2D space indeed bears on the exact motor process responsible for learning.
Project Description:
Sickle Cell Disease (SCD) is a hereditary blood disorder caused by abnormal hemoglobin, which alters the shape and function of red blood cells and impairs oxygen transport throughout the body. This lifelong condition can lead to complications such as chronic anemia, frequent pain episodes, and increased risk of stroke. Non-invasive neuroimaging techniques like transcranial Doppler ultrasound (TCD) and near-infrared spectroscopy (NIRS), commonly used to monitor cerebral oxygenation in SCD patients, may be influenced by factors such as skull thickness.
In this project, Rhea quantified skull thickness in adults with SCD compared to age-matched healthy controls using a novel computational pipeline that extracts and analyzes cranial measurements from MRI scans. She collected and processed MRI data from 30 participants (15 with SCD and 15 Controls) to measure the average thickness of the frontal, temporal, occipital, and parietal bones.
Image preprocessing included bias correction and spatial normalization using the CONN toolbox to standardize image dimensions and coordinate systems. Skull extraction was performed with BrainSuite, an open source software that generates 3D meshes of the inner and outer skull surfaces from MRI inputs. Manual corrections were then made using 3D Slicer to refine segmentations. A custom Python script computed skull thickness by converting skull surfaces into point clouds and calculating the distance between corresponding points on the inner and outer surfaces. Region specific thickness was determined by filtering these measurements according to anatomical landmarks.
The study found that adults with SCD had significantly thicker frontal and parietal bones compared to controls. Future work will focus on optimizing the current segmentation workflow to reduce manual adjustments and expanding the sample size to validate these preliminary findings.
Project Description:
Brain-computer interfaces (BCIs) enable patients with motor-related disabilities to interact with their surroundings using only their brain. The Stentrode is a less invasive alternative to traditional intracranial BCIs that is implanted within the superior sagittal sinus using an endovascular approach. It enables bilateral recording of neural activity associated with movement from the primary motor cortex. However, to realize its potential as an assistive technology, it is essential to characterize the parameters necessary for accurate decoding of motor intent. Previous work explored linear decoding strategies; in this analysis, we compare the classification accuracy of nonlinear shallow learning models such as support vector machine (SVM), k-nearest neighbors (k-NN), and random forest (RF) to a deep learning approach using a convolutional neural network (CNN). Additionally, we aim to identify the point at which adding more electrodes no longer yields significant performance gains for motor classification.
Here, we present data from one participant from an Early Feasibility clinical trial in the United States (ClinicalTrials.gov, NCT05035823). A patient with severe paralysis was prompted to attempt movement of either their left ankle, right ankle, or both ankles. To train the shallow learning models, we calculated features by bandpass filtering the signal to extract neural activity in the high- (70-200 Hz) and low-gamma (30-70 Hz) range then computed the mean amplitude of the signal in 500-ms windows. To train the CNN, entire segments of high- and low-gamma data were used as inputs without windowing or feature extraction. All models were trained on 80% of the data, then tested on the remaining 20% of the data. The accuracy of each model was compared across all combinations of channels from 1-13 channels.
We determined that k-NN and RF models achieve high accuracy at all channel combinations, outperforming both polynomial SVM and CNN models, and that k-NN classification accuracy does not improve with the inclusion of additional channels beyond 6 electrodes. Future work will involve determining which machine learning models can achieve high-accuracy classification with input from fewer electrode channels.