Measuring cardiac stroke volume through in-ear audio sensing

Figures

• 

  Fig. 1. Signal properties of the in-ear audio, ECG, stroke volume signals, and correlations between the in-ear audio signal and stroke volume. a–d 60 s of in-ear audio, ECG, inter-beat interval (IBI) and stroke volume for two participants before and after isometric exercise. a, b Participant 1 before and after exercise, respectively. c, d Participant 2 before and after exercise, respectively. In-ear audio exhibits consistent heart signals, with a peak corresponding to each heartbeat in the ECG signal. Before exercise, stroke volume maintains a consistent level. In the after-exercise condition, stroke volume is elevated and slowly returns to baseline levels. e Correlations between features extracted from the in-ear audio and the stroke volume. Weak correlations exist between heart-related properties and stroke volume, and between audio features and stroke volume. This highlights the potential of using in-ear audio to estimate stroke volume and emphasises the need for advanced learning techniques to intelligently leverage the weak correlations.

• 

  Fig. 2. Overall stroke volume estimation results. a Scatter plot comparing the overall calculated stroke volume (SVpred) and the ground truth stroke volume (SVGT) per user. The Pearson correlation coefficient r quantifies the quality of correlation. b Relationship between the calculated stroke volume and the ground truth stroke volume for the two exercise conditions: before exercise (green circles) and after exercise (blue triangles). c Bland–Altman plot indicating the agreement of the overall estimated stroke volume with the overall ground truth stroke volume per participant. The horizontal lines indicate the mean value of the difference and the 95% limits of agreement (d and d ± 1.96σ). d Bland–Altman plot showing the agreement between the calculated and ground truth stroke volumes over the exercise conditions for each participant. Markers correspond to the two exercise conditions: before exercise (green circles) and after exercise (blue triangles). e Per-participant estimation mean absolute percentage error (MAPE) for the two conditions.

• 

  Fig. 3. Exploded schematic of the custom earable system for in-ear audio collection. a Custom 3D printed earhook based on the design of Röddiger et al.48. b Speaker (a standard true wireless stereo 8mm, 16 Ohm earbud speaker) c Knowles SPU1410LR5H-QB microphone (d) 3D printed front cap (e) rubber ear tip (f) wire (g) custom PCB containing one MCP6004 non-inverting operational amplifier per channel (earbud). These amplifiers are controlled by potentiometers, allowing for adjustable gain if required. The gain value was set to unity for all data collection exercises. The PCB was designed to interface with the audio codec and relays the amplified microphone signals to the analog inputs of the codec. h HiFiBerry DAC+ ADC pro audio codec (i) Raspberry Pi 4. Figure of the audio codec from HiFiBerry 3D models of HiFiBerry boards (HiFiBerry), and model of the Raspberry Pi 4 from GrabCAD (Raspberry Pi 4 Model B, user-uploaded CAD model).

• 

  Fig. 4. System processing pipeline. a Data collection protocol whereby data is simultaneously collected from our earable prototype and the ground truth using the Finapres NOVA continuous stroke volume measurement device. Not pictured is the isometric exercise where the participant squeezes a stress ball. b The pre-processing pipeline for the audio signals involves downsampling, segmenting the audio into 10-second windows with 8-second overlap, and generating a spectrogram for each window. c An unlabelled dataset of in-ear audio is used to pre-train the encoder-decoder network with masked spectrograms, employing a random, unstructured masking strategy. d The pretrained encoder is fine-tuned using the stroke volume in-ear audio dataset to predict stroke volume per 10-second window. During fine-tuning, a structured 2D masking strategy is used. Icon in (c): Internal Medicine, The Noun Project. Licensed under CC0 1.0 (public domain). Source: (Wikimedia Commons). Icon in (d): Cardiology, The Noun Project. Licensed under CC0 1.0 (public domain). Source: (Wikimedia Commons).

• 

  Fig. 5. Overview of the impact of exercise on the participants’ average stroke volume. a Breakdown of the exercise frequency of the surveyed participants. b Comparison of the average stroke volume per exercise frequency (n = 31 participants). In the plot, each box spans from the first to the third quartile, with the horizontal line indicating the median. The whiskers extend up to 1.5 times the interquartile range beyond the box’s lower and upper bounds. c Distribution of beat-by-beat stroke volumes in the collected data for five minutes of the baseline state (pre-exercise) and five minutes after exercise.