The Bike Fit Study
Fifty-one riders, three sessions each, asking whether a professional bike fit actually changes what the body does, and whether it differs between men and women. I designed the protocol, ran every session, and built the pipeline from raw sensor stream to finding.
What it had to do
Professional bike fitting is a large industry built on a fairly small evidence base, and almost none of that evidence addresses whether the effects differ by sex. The study asked a narrow version of the question. After a professional fit, does anything measurable change in physiology, biomechanics, muscle activation or comfort, and does it change differently for women than for men?
It earned 82%, and I am working the study up into a paper over the coming months. The findings stay off this page until that is done, so what follows is the method and what it cost to run.
What I chose
A crossover design. Every participant rides both their own self-selected fit and a professionally adjusted one, so each rider is their own control, which matters enormously when you have around fifty people and enormous variation between them.
Three sessions each. Screening, anthropometrics, self-selected fit and a ramp test to establish FTP. Then a twenty-minute trial at 95% of FTP on the self-selected fit. Then the same trial on the professional fit, with perceived exertion recorded at four, nine, fourteen and nineteen minutes.
To be clear about credit: the measurement platform was custom built by Bath's Lovelace and IMPACT labs. I did not build the rig. My work was the experiment design, running every session, and everything that happened to the data afterwards.
How it works
Each session produces breath-by-breath metabolic data from a COSMED cart, markerless motion capture through Qualisys and Theia3D, EMG and IMU traces, muscle oxygenation from Moxy sensors, plus pressure, strain and respiration streams. A single session's raw JSON runs to one or two gigabytes.
The pipeline reconciles all of it. Different instruments, different clocks, different sampling rates and different file formats have to be aligned to a common timeline before any of it means anything. That synchronisation problem was the real engineering work, and it is where most of the time went.
What it cost
Time, mostly in the lab. Fifty-one participants across three sessions each is over 150 sessions, every one of which had to be run in person and to the same protocol.
The recurring cost was data integrity. Instruments drop frames, sensors come loose, and participants cannot always hold the target power. The protocol steps power down by 5% when that happens, which then has to be modelled rather than ignored. A meaningful share of the pipeline exists to catch and handle those cases, not to compute anything.
What I'd change
I would validate the synchronisation between instruments on day one with a deliberate test session, instead of discovering alignment problems while processing real participants. Some of that data was harder to rescue than it needed to be.