Use non-invasive exercise signals to estimate blood lactate, a useful proxy for metabolic stress and threshold behavior.
Physiological machine learning
Blood Lactate Estimation
Ensemble and Bayesian models estimating blood lactate from wearable physiological signals.
Background
Athletes and clinicians need lower-friction ways to infer metabolic state without relying on frequent invasive measurements.
Blood lactate is a useful marker of metabolic strain, but direct measurement is invasive and sparse. This project explores whether wearable and exercise-test signals can estimate lactate continuously enough to support training or clinical interpretation.
The page emphasizes repeated-measure structure: testing days differ, rest-like zero-power intervals matter, and nonlinear models can capture physiology that a simple linear baseline misses.
Project Workflow
Treat each testing day separately first because baseline fatigue, protocol structure, and sensor behavior can shift over time.
Zero power or cadence often indicates rest rather than missingness, so these rows are surfaced instead of blindly dropped.
Use ensemble regressors to capture nonlinear relationships, then layer Bayesian modeling on top to reason about repeated-measure uncertainty.
Model Comparison
Flexible models improve over the linear baseline.
The model comparison gives a compact view of predictive lift.
Day Summary
Day-level summaries reveal protocol and physiology shifts.
These summaries are useful before modeling because each day has different mean lactate, maximum lactate, and rest-like interval structure.
| Day | Rows | Mean lactate | Max lactate | Rest-like rows |
|---|---|---|---|---|
| Day 1 | 3238 | 4.59 | 9.79 | 833 |
| Day 2 | 3180 | 4.4 | 7.63 | 506 |
| Day 3 | 2400 | 7.56 | 11.16 | 496 |
Associated Presentation
CS 1090A Blood Lactate Final Project
Final project deck for the blood lactate estimation analysis, including data exploration, modeling, and results.
