Biomechanical Signal Analysis & Clinical Data Pipelines
Automated pipelines that transform weeks of raw prosthetic and orthotic sensor data into structured activity timelines and clinical outcomes for real-world ambulatory studies.
Overview
The Problem
Smart prosthetic sockets and orthoses with embedded sensors and microcontrollers generate enormous amounts of continuous multi-sensor data across both in-lab sessions and multi-week take-home studies. The challenge is turning that raw data into something meaningful fast enough to actually drive research and clinical decisions.
The Solution
Built automated pipelines that ingest raw sensor recordings from entire study periods and output structured activity timelines, wear-time metrics, and comparison reports across the full dataset, removing the manual processing burden that would otherwise make this scale of analysis infeasible.
Key Outcomes
- ●Enabled cross-device, cross-participant data analysis and insight that prosthetists and clinicians have never had access to at this scale
- ●Built end-to-end: raw sensor data to meaningful statistical summaries in minutes
- ●Robust software utilizing highly specialized detection algorithms for reliable activity classification across diverse users and device configurations
My Role
Biomechanical Research Engineer II
Gallery
echoRefine pipeline demo: data upload through activity extraction and statistical output
Pin sensor vs. standard sensor analysis comparison for a few seconds of data
Activity classification output comparison between sensor configurations
Day of adaptive socket control + activity data for in-depth analysis
Challenges & Solutions
Signal Variability Across Devices, Users, and Activities
Inter-subject variability in sensor signals is significant. Signal characteristics shift depending on the device, the individual user, and the activity being performed. Each of those dimensions introduces its own patterns and edge cases that have to be individually investigated and accounted for before activity recognition can be reliably automated at scale.
Multi-Characteristic Detection Strategy
Rather than chasing a single better threshold, I systematically identified several distinct signal features through careful feature extraction, then sequenced them so ambiguous cases get resolved rather than collapsed into false positives. The hard part was not the code. It was figuring out which features of the signal were actually meaningful and generalizable across participants.
Processing Efficiency in a Small Lab
With 10+ participants, months of continuous 32Hz multi-sensor recordings, and a small research team, the time cost of manually processing data adds up fast. Hours spent wrangling files are hours not spent on research. Building tools that eliminate that overhead is not a nice-to-have in a setting like this, it is essential.
Automated Cross-Participant Pipeline
Built the system so a researcher can upload data files and get structured, statistically analyzed outputs across the full dataset without manual cross-referencing. The pipeline handles the bookkeeping and the outputs are designed to surface the differences that matter, not just dump raw numbers.
Impact & Results
What's most powerful about this work isn't the numbers. It's being able to sit down with our research prosthetist, look at actual outputs from real participants, and have a productive conversation about how our devices are able to help people, improve further, and inform their care providers for real-world impact. The data is full of biomechanical, human nuance. Learning to read it and build tools that reliably extract meaning from it has been one of the most in-depth and rewarding things I've done. The pipelines continue to support ongoing take-home studies and feed directly into both clinical reporting and the adaptive socket's control logic.
Key Takeaways
- ●Reading a biomechanical signal well enough to build reliable software around it requires genuine domain understanding. You cannot abstract away the physiology.
- ●The most valuable part of algorithm work is not the final accuracy number. It is designing a system that lets you iterate fast enough to actually find out what works.
- ●Clinical context shapes every technical decision. Building for a research prosthetist means the outputs have to be interpretable, not just correct.
- ●Translating deep domain knowledge into reliable, automated software requires genuinely understanding the data at a level most engineers never need to. The analysis informs the architecture, and the architecture has to earn the trust of the clinicians using its outputs.
Skills Developed
Interested in working together?