Interval Lab

A Streamlit app that analyses the speed trace of interval sessions recorded on a GPS watch (Garmin .FIT files). It answers three questions about a set of reps:

1. How was each rep run? For a single rep it finds the moment the runner stopped accelerating and started holding speed (a two-segment linear fit), and reports peak speed, average speed and peak acceleration.
2. Where in the rep do two groups of reps differ? Compare, for example, the first three reps with the last three. Statistical parametric mapping (SPM) tests the whole speed curve rather than a single summary number. Choose a two-sample t-test (Welch) or a non-parametric permutation test; the effect size (Cohen's d) is plotted across the rep.
3. Where in the rep does fatigue show? A regression of speed on rep number at every point in the rep shows how much speed is lost per rep, and where that trend is statistically significant.

Fatigue Trend tab on the synthetic session in sample-data/. Top: each rep (grey) and the mean (black). Bottom: change in speed per rep at each moment of the rep; shaded spans are statistically significant (p < 0.05).

Run it

pip install -r requirements.txt
streamlit run dashboard.py

Upload your own .FIT file, or try sample-data/synthetic-intervals.fit.

Tested with Python 3.14, Streamlit 1.64, spm1d 0.4.54, pandas 3.0 and NumPy 2.5 (September 2026).

How it works

• Reads record messages (timestamp and enhanced_speed) and lap messages. Each lap is treated as one rep, so if your watch records recoveries as laps, select only the work laps.
• Resamples speed to 10 Hz with cubic interpolation, and derives acceleration from the change in speed, smoothed over 0.5 s.
• For the group and trend tabs, each rep is trimmed to the analysis window and time-normalised to 100 points. Set the window no longer than your shortest rep: reps shorter than the window are dropped and listed under Warnings.
• SPM uses spm1d: ttest2 with unequal variances, or nonparam.ttest2 for the permutation test, and regress for the fatigue trend.

Checking it against a known answer

sample-data/synthetic-intervals.fit is simulated, not recorded: 8 reps of 45 s with a fatigue trend built in. Peak speed falls by 0.05 m/s per rep, and after 15 s each rep fades a little more than the one before. Between 22 s and 40 s into the rep, the built-in decline averages 0.41 km/h per rep. The Fatigue Trend tab flags a significant decline from 22 s to the end of the window and recovers 0.41 km/h per rep (shown as 0.4 in the app).

sample-data/make_synthetic_fit.py regenerates the file (it needs pip install fit-tool).

Reading the results carefully

• Reps from one session are not independent samples: they come from one athlete on one day. Treat the p-values as a guide to where differences sit within that session, not as evidence that generalises to other athletes or sessions.
• Watch GPS speed is usually logged about once a second. Resampling to 10 Hz aligns and smooths the trace but adds no information, and acceleration derived from GPS speed is noisy.
• A permutation test needs at least 3 reps per group to reach p < 0.05 (20 possible permutations). With fewer, the app says so instead of running it.

Author

Alan Ruddock, exercise physiologist, Sheffield Hallam University (alanruddock.com).