Predicting physical activity energy expenditure, types, and intensity using long short-term memory networks in aged 6-18 cross-stage populations.

BACKGROUND

Accurately estimating physical activity (PA) and energy expenditure of children and adolescents cross-stage are key issues in sports medicine and beyond.

OBJECTIVE

This study aims to develop and test a long short-term memory (LSTM) network to predict physical activity energy expenditure (PAEE), types, and intensity using accelerometer data collected in children and adolescents.

METHODS

A total of 120 healthy participants aged 6-18 years completed 13 PA tasks while wearing triaxial accelerometers at 9 body sites and simultaneously wore the portable metabolic analyzer. 59 features were extracted from accelerometer signals, combined with demographic data to build LSTM models for cross-stage and single-stage (childhood, early and late adolescence), respectively, and compared with the results of 3 ensemble learning (random forest, extreme gradient boosting, and light gradient boosting machine).

RESULTS

The RMSE for predicting PAEE across and single stages were about 0.55 metabolic equivalents (METs) to 0.73 METs, and an average reduction of about 55% in RMSE over ensemble learning. The RMSE for predicting energy expenditure of a specific PA type was averaged about 0.38 METs to 0.49 METs, and an average reduction of about 60% in RMSE over ensemble learning. The recall for predicting PA types was about 90% to 93%, with an average relative improvement of about 70% over ensemble learning. The recall for predicted PA intensity was about 90% to 93%, with an average relative improvement of about 32% over ensemble learning.

CONCLUSIONS

LSTM showed strong generalizability and robustness in cross-stage PA monitoring, offering an accurate and scalable solution.