People attach greater importance to the physical health of teenagers because adolescence is a critical period for the healthy development of the human body. With the progress of biosensing technologies and artificial intelligence, it is feasible to apply wearable devices to continuously record teenagers' physiological signals and make analyses based on modern advanced methods. To solve the challenge that traditional methods of monitoring teenagers' physical fitness lack accurate computational models and in-depth data analyses, we propose a novel evaluation model for predicting the physical fitness of teenagers. First, we collected 1024 teenagers' PPGs under the guidance of the proposed three-stage running paradigm. Next, we applied the median filter and wavelet transform to denoise the original signals and obtain HR and SpO2. Then, we used the Pearson correlation coefficient method to finalize the feature set, based on the extracted nine physical features. Finally, we built a 1D-CNN with LSTM model to classify teenagers' physical fitness condition into four levels: excellent, good, medium, and poor, with an accuracy of 98.27% for boys' physical fitness prediction, and 99.26% for girls' physical fitness prediction. The experimental results provide evidence supporting the feasibility of predicting teenagers' physical fitness levels by their running PPG recordings.
Keywords: Convolutional Neural Network (CNN); Long Short-Term Memory (LSTM); Pearson correlation coefficient (PCC); Photoplethysmography (PPG); deep learning; noninvasive biosensors; teenager physical fitness monitoring; wearable bracelets; wireless biosensors.