Predicting Marathon Time Using Wearable and Physiological Data

Although training and physiological data from wearable devices are becoming more advanced, interpreting and translating data into realistic goals for marathon runs are still difficult. The current research study devised a method for predicting marathon run times using Garmin training data and physiological data on runners. This method aims to improve the accuracy and safety of runners’ pace selections and mitigate the risk of runners’ extreme fatigue, also known as wall-hit, or exercise-associated collapse, during the marathon. The research study underwent the lifecycle of a machine learning process from data engineering, model engineering, to deployment. The researchers obtained 100 aggregated samples from 16 participants’ Garmin (CSV) files and personal data (age, weight, and gender). The data was cleaned, while transformations on the data, aggregation as well as selection of the features from the data were done. The Multiple Linear Regression model was subjected to evaluation using 10-fold cross-validation as a model. The model produced the best results with an average R-squared of 0.8255 ± 0.0811, RMSE of 538.77 seconds (8.9 minutes), and MAE (Mean Absolute Error) of 435.17 seconds (7.2 minutes). A separate comparison experiment showed that the model that lacked physiological factors in the model performed poorly. The model with physiological factors had its R-squared improved from 0.69 to 0.82, and the MAE improved by 83.20 seconds. A Streamlit prototype finished time prediction model and pacing guidance model was built to conduct the user evaluation. The user evaluation showed that out of the 53 respondents, most of them (92.5%) agreed that the pacing recommendations were useful. These results are evidence for the fact that there exists a practical support for the recreational marathon planners. The support is attributed to the combination of physiological indicators obtained from the wearable devices and regression.