Novel Signal Noise Reduction Method through Cluster Analysis, Applied to Photoplethysmography

Figure 1. Three recorded photoplethysmography (PPG) traces measured from the great toe site. Upper: a clean PPG signal; middle: a signal dominated by low-frequency noise, usually caused by movement of the subject's limb; lower: a PPG trace contaminated with high-frequency noise, typical of electrical interference. Typically, these noise features can appear intermittently within a recording made over a measurement period of 1-2 minutes.

Figure 2. Algorithm's steps represented as a flow diagram, utilising BPMN Notation [16].

Figure 3. The (a) high-pass and (b) low-pass filters. The phase response for the high-pass filter is shown as a dashed red line, showing an increasing nonlinear effect at very low frequencies. The phase for the low pass is linear, such that all frequencies are delayed by the same time.

Figure 4. (a) PPG trace with artificial “movement” noise added. (b) and (c) show full pulse set and clustered pulses. (d) shows full trace with cluster pulses highlighted.

Figure 5. Median peaks from the clean data pulses, all pulses after noise is added, and the median from using clustering on the noisy data.

Figure 6. Comparison between clustering and nonclustering for a variety of inputs. (a) Simulated movement artefacts, (b) simulated electrical noise, (c) simulated electrical artefact and movement noise, (d) no noise, and (e) white noise as an input. Note that in (c) and (e) a cluster of sufficient size was not formed; therefore the algorithm reported no result.

Figure 7. Comparison between the cluster and noncluster algorithms with increasing noise.