Application of a Recurrent Neural Network to Predict the Oxygenated Recovery State Following Maximum Isometric Hand Gripping Exercise

The rest period in between strength exercises determines how the short-term energy supplies in the muscles are replenished and metabolites are cleared. Near-InfraRed Spectroscopy (NIRS) is a proven method to study oxidative recovery kinetics following exercise. The goal of this study is to develop a model that predicts the oxygenated recovery state, this can help athletes optimize the resumption of exercise. 17 healthy subjects performed a sustained isometric hold in a hand gripper until volitional exertion, Tissue Saturation Index (TSI) was continuously monitored throughout and following exercise by a NIRS sensor (Train.Red PLUS). The oxygenated recovery state was manually categorized by three independent experts into four different phases of recovery; I - a pronounced increase, II - a gentle increase, III - the maximum oxygenated state, and IV - the return to baseline. A Recurrent Neural Network, inspired by Natural Language Processing, was trained and tested on this data, resulting in a model that predicts shifts between phases of recovery. A 5-fold cross-validation analysis resulted in the following average performance: • Recurrent Neural Network: Accuracy: 55.17%, categorical cross-entropy: 1.02351. • Multi-Layer Perceptron: Accuracy: 57.16%, categorical cross-entropy: 0.95201. • XGBoost: accuracy: 44.85%, categorical cross-entropy: 10.1119. In predicting the user’s current state of oxygenated recovery the MLP and RNN are similar in performance, however, the MLP shows erratic behavior, while the RNN generally follows the shift in phases of the ground truth. These capabilities could enable athletes with different fitness goals to design goal-tailored and therefore more efficient training.