Jillian Wojcik

Wearable instrumentation · senior design

Wearable Vocal Dosimeter & Activity Monitor

A wearable that records voice use and body movement together. I manage the team, set the requirements and write the firmware. Early bench result: measured speaking time within 2.9% of actual.

Clinicians studying pelvic-floor dysfunction in women suspect that vocal effort and physical exertion interact, but the two are almost never measured on the same clock. I lead a seven-person student team building a wearable that records both at once: voice from a microphone at the neck, body movement from an accelerometer, logged together on an ESP32 to a microSD card with shared timestamps so the two streams can be compared second by second.

As project manager I turned the research question into device requirements, set the targets the hardware has to hit, and coordinate the team against them. I also write and debug Arduino (C++) firmware for sensor acquisition and logging, and build the MATLAB analysis that checks whether the device's voice detection matches what actually happened.

Project status

1. RequirementsResearch question turned into device targets
   Done
2. First prototypeESP32 firmware, neck mic, microSD logging
   Done
3. Bench testingChecking voice detection against scripted trials
   Now
4. Wearable buildEnclosure and synced activity sensing
   Next
5. Wear validationFull-day recordings
   Planned

System design

Teal blocks are running on the current bench prototype. The others are designed but not yet integrated and tested.

Early bench test results

The first test followed a 60-second script: three 10-second voicing periods separated by silence. I compared two ways of deciding "voice on": a simple loudness (RMS) threshold, and a check on whether a stable pitch is present. Pitch held near 120 Hz throughout each voiced period, and the device's running phonation total landed close to the 30 seconds actually voiced. The plot also showed the loudness-only detector firing on brief noise during silence, which is why the pitch check is being kept in the design.

Biomechanics lab · BME 429

Mechanical Testing of Teflon, Rubber and Beef Muscle Tissue

Measured strength, stiffness and fatigue of materials with tensile testers, then compared to a COMSOL model. Result: muscle was about 28× stiffer cut across the fibers than along them.

Before a material goes into a device or a tissue model, you need to know how stiff it is, where it yields, and how it behaves when it's loaded thousands of times. We pulled Teflon dog-bone samples and beef tissue to failure on a Lloyd tensile tester, cycled Teflon and rubber bands 500 times at 1, 10 and 50 Hz on a Bose ElectroForce, and built a COMSOL model of the dog-bone sample to see where stress concentrates.

Beef tissue: stiffness with vs. against the muscle fibers

Beef cut against the grain had a mean modulus of 0.486 MPa. Cut with the grain, it was 0.0173 MPa. A two-sample t-test confirmed the difference. This was the opposite of what we predicted, and it showed how much sample preparation and clamping affect soft-tissue test results.

Cyclic (fatigue) testing results

At 1 Hz, Teflon settled from about 13.1 N to 12.5 N of load amplitude over the first 150 cycles and then held steady, a 4.6% drop. Rubber bands barely changed at any frequency. At 10 Hz, Teflon's stress lagged its strain by roughly 108°, a clear sign of viscoelastic, rate-dependent behavior. At 50 Hz the Bose system captured only about two points per cycle, so I flagged those loops and phase plots as unusable instead of reporting an energy loss the data couldn't support.

Biomaterials lab · BME 429

Blood Cell Analysis and Biomaterial Surface Testing

Counted and identified blood cells under the microscope, and tested how blood fluids spread on glass, Teflon and cell-culture plastic. Result: the surface material, not the fluid, changed how liquids spread (p < 0.05).

Anything implanted in the bloodstream is judged first by its surface: how plasma proteins stick to it shapes clotting and the body's response. We characterized bovine blood cells, measured how plasma, red blood cells and water spread on three common lab surfaces, and tested glycerin–water mixtures as stand-ins for blood viscosity.

Contact angle results

I photographed 15 µL droplets and measured contact angles in ImageJ, nine droplets per fluid–surface pair, then ran one-way ANOVAs both ways. Surface type changed the angle significantly for every fluid (all p < 0.05). Fluid type made no significant difference on any surface. Teflon sat right around the 90° line where liquids stop wetting a surface.

Research experience

Ultrasound and Photoacoustic Imaging for Prostate Cancer

Undergraduate researcher on an NCI-funded imaging project: equipment setup, sample prep and cell culture. The group published the system in Proceedings of SPIE.

Photoacoustic imaging fires light into tissue and listens to the ultrasound it generates, which reveals vasculature and oxygenation that conventional ultrasound misses. It also means the data is only as good as the setup: small inconsistencies in alignment, coupling or sample preparation show up as artifacts you cannot remove later.

My contribution was to that reproducibility. I configured and troubleshot imaging equipment, prepared biological samples and experimental materials, and standardized acquisition setups so runs could be compared to each other. I also maintained, passaged and expanded mammalian cell cultures four to eight times a week, checking morphology, confluency and density under the microscope before imaging time was spent on them.

Computational modeling · COMSOL

Coronary Stent Design Comparison Using Blood Flow Simulation

Simulated blood flow through a coronary artery with four different stent designs to find which one disrupts flow least. Result: the thin elliptical-strut design was best.

Stent struts sit directly in the bloodstream. Their shape changes the velocity field and the shear the artery wall feels, and abnormal wall shear stress is associated with endothelial damage and platelet activation. Holding hole size constant, we asked whether strut cross-section and thickness measurably change hemodynamics in a left anterior descending bifurcation.

I built the arterial model and researched the clinical scenario it had to represent. We subtracted four stent designs (elliptical and rectangular struts at 0.2 mm and 0.4 mm) from the lumen and solved steady laminar flow against an unstented control: inlet 0.2 m/s, outlet 13,332 Pa, blood at 1060 kg/m³ and 0.0035 Pa·s. Reynolds number 187 and Womersley number 2 justified the steady laminar treatment.

Downstream velocity recovery as a fraction of inlet velocity, averaged over seven cut points. Thickness dominates.

Wall shear stress results

Velocity recovery alone would have picked the rectangular 0.2 mm design. Wall shear stress reversed that: the elliptical 0.2 mm design had the lowest peak stress and by far the smallest spread between maximum and minimum.

Analog signal conditioning

Analog EEG Filter Circuit for Four Brainwave Bands

Designed and built a circuit that splits a brain signal into delta, theta, alpha and beta waves and amplifies each. Result: all four bands met spec with under 2% error, for $9.60.

A scalp EEG arrives as one crowded, microvolt-scale trace. Clinically, what matters is how much energy sits in each band: delta for deep sleep, theta for drowsiness, alpha for relaxed wakefulness, beta for active concentration. The brief was to separate all four from a raw input, amplify each to at least 2 V peak-to-peak, and keep the build under $11.

I split the input into four identical parallel branches rather than cascading them, so a failure in one band couldn't corrupt the others. Each branch is a high-pass RC, a low-pass RC and an LM741 non-inverting gain stage. With a 1 µF capacitor fixed across every stage, each corner frequency comes down to one resistor, and the weakest bands got the most gain.

Signal processing · machine learning

ECG Arrhythmia Classifier in MATLAB

MATLAB program that cleans heart recordings, finds each heartbeat and labels the recording normal or abnormal. Result: 96% accuracy on 200 recordings.

Two hundred ECG recordings, each carrying baseline drift from respiration and electrode movement. I built the full pipeline in MATLAB: a high-pass filter to flatten the baseline, adaptive thresholding to find R-peaks whose amplitude varies between records, and four physiological features from the R–R intervals, heart rate and its variability among them.

A 5-nearest-neighbors classifier on those four features reached 96% cross-validation accuracy separating normal from abnormal traces. The final model then scored 100 unseen recordings; ground-truth labels weren't provided for those, so that run shows throughput rather than a second accuracy figure.

Analog design to hardware

Five-Stage Audio Circuit on Breadboard and PCB

An audio mixer with vocal removal, bass/treble control, volume, an LED level meter and a headphone driver. Result: all five stages worked within 3% of predicted values.

I built the same design twice: once on a breadboard at ±15 V from a myDAQ, then on a printed circuit board running from two 9 V batteries. Dropping the rails changed the comparator thresholds in the LED display, so the resistor network had to be recalculated for the second build.

I simulated all five blocks in Multisim, verified each against the oscilloscope on the breadboard, then soldered the PCB low-profile components first and debugged it stage by stage.

Digital logic

Digital Logic Circuits: Input Detector and 3-Bit Counter

Two hardware builds: a circuit that detects specific 4-input patterns, and a counter with two modes. Result: every input combination tested and correct.

The first build was a four-input identifier, high only for the exact input combinations that mattered. Don't-care conditions collapsed the truth table; I minimized to both sum-of-products and product-of-sums, built both in hardware and confirmed they were equivalent across all 16 inputs.

The second was a 3-bit sequential circuit with a mode input: X = 0 counts 0 through 7 and wraps; X = 1 snaps down to the nearest power of two at or below the current state and holds. Wiring JK flip-flops as T flip-flops turned the excitation table into a question of which bits toggle, which cut the logic to minimize.