The selection of weights precision for ballistocardiography classification

Artificial neural networks have been effectively applied to a variety of medical diagnostic and classification situations. More recently, there is a growing interest in implementing such algorithms for real-time monitoring or fast-time scanning in the classification of normal and abnormal. These applications will benefit from hardware implementation. It is well known that the input and weights precision has a considerable impact on the circuit's complexity, speed and power consumption. This paper evaluates two training methods for selecting limited precision weights for ballistocardiogram classification and suggests that a 7-bit weight precision can provide a similar performance compared with a high precision weights expression. Furthermore, a method is proposed for implementing the artificial neural networks using field programmable gate arrays for the rapid prototyping of algorithm specific hardware.