Non-invasive Lactate Threshold Estimation Using Machine Learning

Abstract

The Lactate threshold (LT) has gained special attention in the sport world and is considered one of the potential indicators to evaluate individual performance in different sports. Traditionally, measuring LT requires frequent collection of blood samples from individuals under specific spatiotemporal conditions. This procedure causes discomfort to individuals besides test related cost. In this paper, we propose a non-invasive model to estimate LT using a machine learning (ML) algorithm as a step towards eliminating the need of blood sample collection and facilitating non-invasive performance test. We train and test this model on a 100-subject dataset, which we constructed in collaboration with Peak Center for Human Performance. We also propose a method to fill the missing values in this dataset, which contains the collected data of real life incremental running tests performed at this Centre. We also shed the light on the correlation between demographic data and the LT occurrence and hence help determine the factors affecting LT estimation as a vital sign in the sport world. Applying a multi-layer perceptron (MLP) algorithm on the constructed dataset provided the best correlation coefficient of 0.7983 compared with the LT ground truth scores. Using different combinations of demographic data in conjunction with heart rate (HR) and speed in the training and testing provided various correlation coefficients, which are also presented in this paper.