Predicting heart rate during physical activities using artificial neural networks

• Autor(a) principal: Davi Pedrosa de Aguiar
• Data de Publicação: 2021
• Tipo de documento: Dissertação
• Idioma: eng
• Título da fonte: Repositório Institucional da UFMG
• Texto Completo: http://hdl.handle.net/1843/45612
• topic: Neural Networks; Heart Rate; Computação – Teses; Redes neurais (Computação) – Teses; Frequência cardíaca – Teses

The heart rate (HR) is an important metric widely used by professionals and amateurs in endurance training as a proxy to physical strain, as it is through heartbeats that oxygen, nutrients and hormones are distributed to cells in the whole body. This metric is important for the prescription of physical exercises, as an effective training should elicit a HR within a certain range, so as to neither under-train, nor over-train an individual. Predicting the heart rate dynamics, nonetheless, is recognized as a hard task, due to a variety of influencing features, ranging from nutrition and mood to an individual’s genetics. Yet, the physical activity is regarded as one of the main drivers of the heart rate. There has been a few studies attempting to model the heart rate using different proxies for physical activities, such as speed and acceleration for running, or torque for cycling. Although these metrics might be good descriptors of the physical strain on the individual, for these specific activities, they are not general enough for describing the strain incurred by other activities, such as rope jumping. Measurements from Inertial Measurement Unit (IMU) sensor, such as accelerometers and gyroscopes present in smartphones and fitness watches, have been successfully applied in predicting the activity being performed by an individual, a task widely known as Human Activity Recognition (HAR). This suggests that these sensors could, in principle, provide more general representations of one’s physical strain, even if it is through the prediction of the activity. Very few published studies attempted using IMU signals to predict the heart rate, and, the ones that did, had some serious limitations, such as predicting over only a few seconds into the future before requiring re-calibration or considering only a single individual, raising some serious questions over its general applicability. In this dissertation, we propose a new model for HR estimation using IMU data, based on Recurrent Neural Networks (RNN). The rationale behind our model is that the same activity elicits different HR responses in different individuals, depending on the physical conditioning. Hence, our model attempts to encode the physical conditioning of an individual into a vector, dubbed PCE, using a specially designed subnetwork. The PCE is then used to initialize the initial hidden vectors of a RNN, that use long short-term memory (LSTM) units. xv We reinforce this encoding by jointly training a network which discriminates whether two PCEs belong to the same individual. We evaluate the proposed model when predicting the HR of 23 subjects performing a variety of physical activities, from IMU data available in public datasets (PAMAP2, PPGDaLiA). For comparison, we use as baselines the only model specifically proposed for this task and two adapted state-of-the-art models for the closely related task of HAR. Our method, named PCE-LSTM, yields over 10% lower mean absolute error. We demonstrate empirically that this error reduction is in part due to the use of the PCE. Last, we use two datasets (PPG-DaLiA, WESAD) to show that PCE-LSTM can also be successfully applied when photoplethysmography (PPG) sensors are available, outperforming the state-of-the-art deep learning baselines by more than 30%.