Explore how wearable sensors capture the body in motion. Control the runner to see how ECG, EMG, PPG, and EEG respond to exercise, and how movement affects signal quality.
These four signals are synthetic and physiologically informed, not human recordings, and time is compressed. They share one model of speed, cadence, gait phase and heart rate. Cadence and motion artifact respond quickly to speed, heart rate lags. Hamstring bursts follow the stride cycle, and artifacts sit at the step rate and its harmonics, as in wearable and mobile EEG recordings. Amplitudes, the EEG rhythms and the exact timings are illustrative. Methods & Validation
Eighteen years of ballet made Evangelia curious about how the brain runs the body.
After years of competitive training, that curiosity led her to neuroscience at the University of Michigan and now a master's in neural engineering at Columbia. Her work sits at the intersection of neural and physiological sensing, signal processing, machine learning and human movement, with experience in EMG, gait analysis and wearable technology.
As a personal trainer, she sees the challenge of translating science into practice firsthand. Advanced sensing technology is still concentrated in research and clinical settings, while consumer wearables are accessible but rely on generalized metrics that miss individual differences and rarely turn measurements into personalized insight.
Her ambition is to bridge that gap with intelligent technology that combines neural and physiological sensing, signal processing and machine learning to understand and ultimately enhance how the brain and body function, communicate and adapt. Because every person has a unique physiological baseline, she envisions technology that goes beyond monitoring to help people understand their own physiology, overcome limitations and expand their capabilities, in health and in performance.
M.S. Biomedical Engineering, Neural Engineering
Expected December 2026
Dean's M.S. Academic Excellence Fellow
B.S. Neuroscience
B.S. Modern Greek Language and Literature
August 2024
University Honors. Carrie Arbour Scholarship. N2N Neuropsychology Prize. Vasilis Lambropoulos Modern Greek Translation Prize.
Robotics and Rehabilitation Laboratory
60-minute cross-training sessions
National Hellenic Student Association of America
Graduate and personal projects are on the .
Methods
Tools
Move your cursor across the stage.
MODEL / SAMPLING / CHECKS / REFERENCES / LIMITS
The four traces on the portfolio home page (ECG, EMG, PPG, EEG) are synthetic. They come from one shared physiological state, so heart rate, gait, muscle activity and motion artifact stay consistent with each other. This page states what is modelled, what was tested, and what is simplified. It is a physiologically informed educational simulation. It has not been validated against experimental human recordings and is not a clinically validated model.
The simulation is an illustration of how movement changes wearable signal quality. It is not a physiological model for estimating anything, and no real recordings are used. Amplitudes are in arbitrary display units.
All four channels are generated at 250 Hz of simulated time and drawn in a 7 s rolling window (1750 samples) with a 1 s grid of 250 samples. A real-time accumulator advances the model in fixed 4 ms steps, so scroll speed and the grid stay tied to simulated time whatever the screen's frame rate. The R-wave phase is stored in a ring buffer so the PPG can look up the cardiac phase from one pulse arrival delay earlier. Random noise uses a seeded generator, so a run is repeatable. In the page, measured output through every stage change was 249.4 to 251.1 samples per second.
These checks test internal consistency (do the channels agree with each other and with the model's own rules) and physiological plausibility (are values in sensible ranges and trends). They are not validation against experimental human recordings. Each result below is computed from samples the simulation generates.
| Check | Result | |
|---|---|---|
| R-R interval equals displayed heart rate at rest, walk, jog, hard | difference under 1.5 bpm; at 139 bpm R-R is 432 ms, at 105 bpm 572 ms | PASS |
| R-R variability falls as intensity rises | rest > walk > hard | PASS |
| ECG peaks coincide with simulated R events | max offset under 12 ms (4 ms sample period) | PASS |
| QRS width and amplitude independent of heart rate | 71 ms and 0.99 at R-R 940, 600, 360 ms | PASS |
| T-wave timing shortens as R-R shortens | T peak 262, 209, 162 ms after R | PASS |
| PPG maximum upslope arrives after the R wave, same rhythm | 150–300 ms (measured to max upslope, which is later than the pulse foot), pulse period equals R-R | PASS |
| EMG main burst precedes ipsilateral heel strike; one complex per stride; quiet at contralateral strike | walk, jog, hard | PASS |
| EMG amplitude rises less than proportionally with speed | hard vs walk under 4× for 4× speed | PASS |
| EMG power is high frequency | 5% below 15 Hz, 95% above | PASS |
| EEG mu peak at rest | 8–13 Hz | PASS |
| EEG gait artifact at stride rate, step rate and harmonics | >1.5× off-harmonic baseline | PASS |
| Artifact-to-neural ratio grows with speed | below 0.35 at rest, above 1 running hard | PASS |
| Cadence and motion artifact respond faster than heart rate | time to 63% of change compared | PASS |
| Heart rate stays elevated after dropping from hard run to jog | more than 10 bpm above steady state | PASS |
| Final walking cadence 100–115 steps/min | within 98–118 | PASS |
| Full sequence returns to REST; no NaN or jumps | REST at about 56 s | PASS |
| Display limiter rarely engaged (no flat tops) | all channels under 2% of samples near limit | PASS |
| No overstriding: foot lands near hip, shank near vertical, knee flexed | foot ≈2 px ahead of hip, shank −3° to −6° | PASS |
| Stance foot does not slide on the ground | difference below 0.05 px/s | PASS |
| Drawn heel strikes per minute equal displayed cadence | 108 vs 109, 174 vs 176 | PASS |
Each entry was opened and read for this page, except where noted. Support is limited to the claim stated.
The checks run the same simulation code as the page headlessly, at 250 Hz of simulated time, and compute each result from the generated samples. Results above come from that run on the current published version.