Session P7A.6

An Artificial Neural Network Model as a Tool to Identify the Anaerobic Threshold during Dynamic Physical Exercise

AC Silva Filho*, RM Souza, L Gallo Jr

Centro Universitário de Franca
Franca, Brazil

Anaerobic threshold (AT) is one of the most important parameters used in exercise physiology. It signals a power value during dynamic physical exercise where anaerobic energy formation for muscle contraction is added to the aerobic counterpart, what allows the quantification of aerobic capacity. In this study, we describe the development and validation of an artificial neural network model to identify AT based on electrocardiogram R-R interval time series collected during physical exercise tests applied in sixteen healthy voluntary persons, of male sex. They presented a sedentary life style and belonged to various age categories: eight young with 25±1.5 (mean ± standard deviation) years and eight of mid age with 42±2.5 years. The electrocardiogram signals (ECG) were initially recorded at rest conditions in two body positions: supine (15 minutes) and seated (15 minutes). Exercise tests were always carried out using a cycloergometer, with the volunteers in the seated position. All the individuals were submitted to three types of tests: a continuous protocol and two others discontinuous protocols (with gradual and alternated increasing powers), each one separated by an interval no more than fifteen days. The continuous protocol was made with gradual increments of 25 Watts (W), at each three minutes, until the physical exhaustion. The R-R series during exercise were obtained from ECG recordings. A feedforward multi-layer perceptron, fully connected, trained with backpropagation algorithm, was used as a tool for the identification of the AT. First order statistics of RR time series, effort power, and 10th order autoregressive model were taken as the network inputs. The results showed that the artificial neural network had its best performance in gradual increasing power. Scatter plot and ROC curve were constructed showing high correlation (r = 0.93), and good accuracy (area under the ROC curve = 0.9851) when compared to Autoregressive Integrated Moving Average (ARIMA) statistical method. An advantage for the use of Neural Nets is the reduced time of necessary computation compared to methods of nonlinear analysis.

(Abstract Control Number: 184)