Heart rate estimation from wrist-type photoplethysmography signals during physical exercise

•A robust and computationally efficient de-noising technique that combines three stages of cascaded adaptive filters like RLS, NLMS and LMS has been proposed.•The filtered output obtained from the RLS, NLMS and LMS adaptive filters are combined using softmax activation function.•The search range (SR), a band of location points from which the HR is estimated, has been selected dynamically by varying the parameter Δ, which depends on the power value of the accelerometer signals.•The proposed method delivers less error value when evaluated on 22 datasets each performing different activity. Wearable devices, such as smart watch use photoplethysmography (PPG) signals for estimating heart rate (HR). The motion artifacts (MA) contained in these PPG signals lead to an erroneous HR estimation. In this manuscript, a new de-noising algorithm has been proposed that uses the combination of cascaded recursive least square (RLS), normalized least mean square (NLMS) and least mean square (LMS) adaptive filters. The MA reduced PPG signals obtained from these cascaded adaptive filters are combined using the softmax activation function. Fast Fourier transform (FFT) is used to estimate the HR from the MA reduced PPG signals and phase vocoder is used to refine the estimated HR. The performance of the proposed method in the form of mean error, standard deviation of the mean error and mean relative error is analyzed using the 22 datasets given for IEEE Signal processing cup 2015. This resulted in an error of 1.86 beat per minute (BPM) tested on 22 datasets which is less compared to other existing methods.