Journal Logo

EP-01 Fitness Assessment, Exercise Training, and Performance of Athletes and Healthy People

Machine Learning Fails To Improve Marathon Time Prediction Compared To Multiple Linear Regression

161

Foreman, Nicholas; Hesse, Anton; Lundstrom, Chris

Author Information
Medicine & Science in Sports & Exercise: August 2021 - Volume 53 - Issue 8S - p 49
doi: 10.1249/01.mss.0000759656.97049.16
  • Free

INTRODUCTION: Predicting sport performance is common for professional athletes but also interests recreational athletes, and previous approaches have primarily used a multiple regression (REG) approach. Newer prediction techniques from machine learning (ML) and more powerful computers may result in more accurate predictions than regression, even for more inexperienced runners.

PURPOSE: To compare REG and ML methods in predicting marathon time using physiological testing variables.

METHODS: Recreationally active college students (n = 127, 88 female) from 2016-2018 performed physiological testing both before and after training for a marathon. Prior to marathon training, participants completed a 2-mile time trial (2MI), hydrostatic weighing for body fat percentage, steady state exercise (a 6-minute run at 75% velocity of 2MI), and a graded exercise test for VO2MAX; these tests were repeated following marathon training. Changes between pre- and post-training variables were expressed as delta values and also included in the dataset. Marathon performance was predicted with REG using principal component analysis (PCA) and backwards elimination, and ML methods included random forest (RF), recursive partitioning (RPART), and bagging (BAG). ML models were run with and without PCA for comparison. Ensembles (ENS) of available regression methods in R’s caret library were also used. Prediction approaches were compared with an out-of-sample root mean squared error (RMSE) on ~20% of the data (n = 28).

RESULTS: The RMSE for the different methods were: REG: 0.440; RPART: 0.509; RF: 0.548; BAG: 0.550; ENS: 0.528. PCA increased RMSE in RPART (1.027), RF (0.677), and BAG (0.705).

CONCLUSIONS: REG outperforms RF, RPART, and BAG when predicting marathon time, and ML prediction worsens when PCA is applied. Ensembles of ML models fail to improve marathon time prediction over individual ML models.

Copyright © 2021 by the American College of Sports Medicine