Expert / Gen1 baseline: 74% accuracy.
Medical devices
FlowSense Clinical / ACE
Clinical wearable and DSP/ML workflow for noninvasive CSF shunt-flow assessment; improved from a 74% baseline to 0.81 AUC / 82% blinded validation accuracy.
Part of Senior R&D Engineer · Rhaeos, Inc. · May 2022 – Dec 2025
completed 2022–2025 Medical devicesAlgorithmsValidationThermal sensing
Type product / algorithm
Skills Wearable sensing, Thermal physiology, ML/DSP, Clinical validation
Selected facts
Quantitative details and source-backed proof points.
FlowSense Gen2 blinded clinical validation: 0.81 AUC / 82% accuracy.
Final FlowSense + imaging framing: 92% sensitivity / 97% NPV.
Multicenter clinical workflow reference: 9 hospitals, 182 subjects, and a 112-subject validation set.
Deck-reported CT/MRI workflow reference: 88.9% sensitivity and 91.0% accuracy.
Supported 3,000+ hours of physiological/device data and 200+ HA Connect datasets across Rhaeos programs.
Workflow included ingestion, provenance, labeling, QC, dataset versioning, feature generation, validation, bias checks, explainability, and deployment checks.
Deployment work included MATLAB-to-TypeScript random-forest conversion with deterministic output checks.
Project summary
Why it exists, what I built, and what I learned.
Why I built it
Clinical shunt-flow assessment needed repeatable model development from noisy thermal physiology instead of manual thresholds alone.
What I built
Automated ingestion, provenance, labeling, QC, dataset versioning, feature generation, validation, bias checks, explainability, and deployment-check workflows.
What worked
The strongest result was connecting sensor physics, clinical usability, small-dataset controls, and deterministic deployment checks into one regulated workflow.
What failed
Small, correlated clinical datasets required patient grouping, leakage checks, bias review, feature-correlation checks, and overfitting controls.
What I learned
Physiological ML is strongest when feature design, validation, and deployment checks stay grounded in sensing physics and clinical workflow constraints.
Stack
Tools, systems, and technical areas involved.
Thermal sensingFeature engineeringMATLABPythonTypeScript deployment checksValidationFDA-ready documentation
Links and direction
Public links and next steps.
Use the same sensor-physics-first validation pattern for adjacent wearable physiological monitoring problems.
Related projects
Other projects in the same neighborhood.
Medical devices
FlowSense Home / Lynx
Home-use hydrocephalus wearable advanced across 3 design generations, 100+ units, 200+ participants, 2,800+ wear hours, and 70% → 96% reliability improvement.
Medical devicesWound Monitoring Platform / Tabby
NIH R43 Phase I multimodal wound patch with thermal/humidity/temperature sensing, 40 mAh battery, 20-subject dataset, 172 logs, and 5-fold wound-model validation.
ImplantablesWireless Battery-Free Bioelectronics
Shared Ph.D. implantable platform work across 14 journal articles, 20+ fixtures, 6 tuned antennas, 3 simulation frameworks, <50 mg implants, and up to 2 m wireless power.