Arash
Keshavarz.
I build machine learning systems for signals from the human body: wearable electrodes, optical pulse, and hyperspectral skin imaging. From the sensor setup to a tracked, deployed model.
Selected work
4 projectsHematoma age prediction
Estimating how old a bruise is from its spectrum
Bruise age is judged by eye today, which is subjective. I built a hyperspectral pipeline that captures skin with a Specim IQ camera at 204 spectral bands and predicts hematoma age from the spectral signature.
- Approach
- Skin segmentation with SAM, PCA and t-SNE, SVM classifier, age regression
- Result
- About 30% lower MAE in age prediction
- Status
- Published, Journal of Imaging, first author
Thigh muscle force estimation with EIT
Estimating muscle force from a soft impedance belt
Measuring muscle force usually needs a dynamometer in the lab. A belt of 16 electrodes around the thigh injects a small current and measures the voltages that change as the muscles contract. I recorded synchronized EIT and torque data from 15 participants and trained Random Forest and SVM models to estimate force levels from those signals.
- Data
- 15 participants, about 80,000 synchronized EIT and torque samples
- Result
- 91% classification accuracy
- Status
- Master's thesis
Hyperspectral SpO2
Pulse and oxygen saturation from a camera
The setup was silently dropping 59% of frames, so the pulse signal was unusable. I traced it to throughput, lowered the frame rate and added spatial binning, then extracted PPG and computed SpO2 with the ratio of ratios method.
- Problem
- 59% frame loss during acquisition
- Result
- Clean PPG at five target wavelengths
- Finding
- Standard calibration constants fail in reflectance mode
medassist-ai
An agent that reasons over wearable health data
Add one sentence on the question a user asks and how the agent answers it.
- Approach
- Model, tools, retrieval, evaluation
- Deployment
- Docker, live demo
- Status
- In active development
Expertise and toolkit
Biosignal ML
Models for EIT, PPG and SpO2 time series from wearable sensors.
PyTorch · CNN · RNNHyperspectral imaging
Acquisition with line scanners, segmentation and spectral regression.
Specim · SAM · YOLOSignal processing
Filtering, Hilbert transform, matched filters, radar fundamentals.
NumPy · SciPy · MATLABMLOps and GenAI
Tracked, reproducible pipelines and LLM agents shipped in containers.
MLflow · DVC · DockerFrom lab hardware to working models.
I like the whole chain: setting up the sensor, fixing the data, training the model and making it run outside a notebook.