Skip to main content
Back to timeline
arXivSource publication:

An AED spots shockable rhythms mid-compression without new sensors: six noise conditions pass the adjudicable AHA criteria at 0 dB artifact

Synopsis

Public defibrillators must pause chest compressions to read the ECG because compressions bury it in artifact; a three-network cascade classifies the contaminated signal directly using only the ECG channel the device already has, with a network biased toward the proposed outcome verifying each decision and disagreement yielding an indeterminate result. On an independent manufacturer-provided cohort of 464 subjects, at 0 dB compression artifact with an eight-second window across six noise conditions, sinus-rhythm specificity was 99.20–99.43% and specificity for other non-shockable rhythms 95.48–96.20%, with only 1.07–1.48% of segments indeterminate; at the harder artifact level the two adjudicable criteria are no longer met.

Source-provided article image: Sensor-Free Shockable Rhythm Classification During CPR Using the Existing ECG Channel of Public-Access Automated External Defibrillators
arXiv

Interpretation

Shockable-rhythm classification can run while chest compressions continue, without adding any auxiliary sensing channel to the defibrillator, using only the single ECG channel the device already acquires. Earlier routes either relied on auxiliary measurements such as transthoracic impedance or compression depth to estimate and subtract the artifact, or reconstructed a cleaner ECG with one network before classifying it; the first needs a sensing channel many existing devices lack, and the second passes reconstruction error to the downstream classifier. On the independent Defibtech cohort of 464 subjects, at 0 dB compression artifact with an eight-second window across six noise conditions, sinus-rhythm specificity was 99.20–99.43% and specificity for other non-shockable rhythms 95.48–96.20%, satisfying both adjudicable American Heart Association criteria, with 1.07–1.48% of segments indeterminate.

The classifier reads compression-contaminated ECG segments directly, with no artifact removal and no reconstruction of a clean waveform. Three identical compact convolutional networks differ only in the class ratios used during training: a balanced network makes the initial call, a shockable-biased network verifies shockable calls, and a non-shockable-biased network verifies non-shockable calls, with disagreement producing an indeterminate outcome. Each decision runs only the balanced network and one biased network and takes 6.68 ms on a single general-purpose processor thread; two single-network baselines matched in parameter count and inference cost do not reproduce this performance, and the widened network failed the sinus-rhythm specificity criterion in all six 0 dB conditions.

Perspective

The results come from synthetic test signals that add clinically recorded compression artifacts to clean ECG, covering a manufacturer cohort of 464 subjects, an eight-second window and six noise conditions at 0 dB compression artifact; the authors state that the next step is evaluation on recordings acquired during actual chest compressions.

The compression contamination is synthetic, so additive mixing cannot fully reproduce the physiological and mechanical interactions present when compressions and cardiac activity are recorded together; at the harder compression-artifact level the two adjudicable specificity criteria are no longer met; asystole has only nine subjects, and its specificity varies by 10.2–13.0 percentage points across partitions, too little to establish a reliable performance level.