Contactless Cardiac Arrest Detection Using Smart Devices

• Agonal breathing is a hypoxia-triggered brainstem reflex that appears in approximately half of cardiac-arrest 9-1-1 reports and is associated with shorter time since arrest and higher survival.Two-thirds of out-of-hospital cardiac arrests in private residences occur at home, where an audible biomarker could be useful.Source: Introduction · p. 1
• Smartphones and smart speakers could identify this audible biomarker and connect unwitnessed cardiac-arrest victims with EMS or people able to administer CPR; smart speakers were projected in 75% of US households by 2020.Source: Introduction · pp. 1–2
• Because agonal breathing is rare, lacks gold-standard measurements, and cannot be reproduced in a lab, the study uses confirmed-arrest 9-1-1 audio and household and sleep-study audio as negatives.It tests whether an SVM can detect agonal breathing in bedroom and sleep settings and classify it in real time on commodity smartphones and smart speakers.Source: Introduction · p. 2

Results

• The SVM classified 2.5-second segments using VGGish embeddings compressed with principal component analysis; evaluation used 82 hours of held-out sleep audio (117,895 segments).Training comprised 7,316 positive and 7,339 negative samples, including agonal-breathing recordings and sleep or domestic-interference sounds.Source: Results · pp. 2–3

Classifier Performance

• AUC was 0.998 ± 0.004, with 97.03% sensitivity and 98.20% specificity; classification took 21 ms on a smartphone and 58 ms on a smart speaker per 2.5-second segment.The operating-point sensitivity and specificity are reported with 95% confidence intervals of 96.62–97.41% and 97.87–98.49%, respectively.Source: Figure 2 · p. 3, Classifier Performance · p. 4
• False positive rate was 0.1009% using classifier probabilities alone, 0.0025% with a frequency filter requiring two breaths in 10–20 seconds, and 0% requiring three.The unfiltered rate corresponds to 119 of 117,985 segments; the filter checks for a rate of 3 to 6 agonal breaths per minute.Source: Figure 2 · p. 3, Classifier Performance · p. 4

Benchmark Performance

• At distances up to 3 m, both devices exceeded 98.7% mean accuracy; accuracy averaged 96.95% across interference classes, while cancellation raised soft/loud self-audio accuracy from 11.84%/3.82% to 97.60%/95.41%.With the smartphone in a pocket and the subject supine, mean detection accuracy was 92.31% ± 6.11%.Source: Benchmark Performance · p. 4

• Passive agonal-breathing detection could reach previously unreachable residential cardiac-arrest victims and may also benefit elder-care settings, EMS dispatch, and overdose-risk monitoring.Source: Discussion · p. 5
• Because an unconscious person cannot intentionally activate a device in an emergency, the envisioned system would run locally and store no data to address privacy concerns.Source: Discussion · p. 5
• Approximately 50% of cardiac-arrest victims have agonal breathing, leaving others undetected; estimates based on 9-1-1 calls may underestimate its incidence.Contactless detection avoids requiring a victim to wear a device and can use existing wired smart speakers without power constraints.Source: Discussion · pp. 5–6
• The study used agonal instances from one community over eight years, and further data are needed to test generalization and distinguish agonals from other acute breathing disorders.Potentially confounding conditions include seizure, hypoglycemia, severe stroke, and drug overdoses with disordered breathing; these were not established as distinct from or similar to agonals.Source: Discussion · p. 6
• Real-world deployment would require an alarm and a cancellation interface before emergency services are activated, since this proof-of-concept study did not activate EMS.Source: Discussion · p. 6
• Growing commodity smart-speaker adoption in private residences and hospitals may provide a wide-reaching path for contactless cardiac-arrest detection.The discussion also notes that technology evolution may create opportunities to improve health.Source: Discussion · p. 6

Datasets.

• Researchers annotated 236 agonal-breathing clips from 162 high-confidence 9-1-1 calls and assembled 83 hours of sleep-lab negatives from 12 patients.The calls came from 729 recordings totaling 82 hours; the sleep data included apnea, hypopnea, snoring, breathing, and interfering sounds. The study was IRB-approved, and call data were provided by Public Health–Seattle & King County EMS.Source: Datasets · p. 6, Materials and methods · p. 6
• The team augmented recordings across distances, devices, noise conditions, and phone connections, yielding 14,655 cross-validation examples.The dataset contained 7,316 agonal-breathing and 7,339 negative instances.Source: Datasets · pp. 6–7

Data preparation.

• Audio was normalized and passed through VGGish, which resampled it to 16 kHz and produced 256-dimensional embeddings from log-mel spectrograms.Input clips were sampled at 8 kHz; the spectrogram was computed using the Short-Time Fourier Transform, then PCA was applied.Source: Data preparation · p. 7

Training algorithm.

• The classifier used an RBF-kernel SVM with C=10 and 10-fold validation that kept breathing instances, recording conditions, and calls separated across splits.Validation files were excluded from training under other recording conditions, and recordings from the same call did not straddle the split.Source: Training algorithm · p. 7

Benchmark experiments.

• Benchmark experiments recorded 236 clips on an Echo and iPhone 5s across distances, while separate tests assessed music interference cancellation and sounds played outside the room.The interference-cancellation test used music at 45 and 67 dBA and a 100-weight Sign-Data LMS filter with step size 0.05; external interfering sounds were played at approximately 55 dBA at 1 meter.Source: Benchmark experiments · p. 7

Run-time Analysis.

• On an iPhone 7, spectrogram generation took 16 ms and SVM inference 5 ms; on an iPhone 4 used to estimate Echo runtime, the respective times were 40 ms and 18 ms.The iPhone 7 implementation used Accelerate for FFTs and Monte Carlo sampling to approximate the RBF kernel; third-party local analysis was not then allowed on Echo devices.Source: Run-time Analysis · p. 8

• Table S1 summarizes sleep-study patient demographics, hypopnea and apnea event counts, and the apneas-hypopneas index (AHI).Source: Table S1
• Table S2 summarizes the demographics of people with agonal breathing during 9-1-1 calls.Source: Table S2

• Supplementary Table 1 summarizes sleep-study patients’ demographics, hypopnea and apnea event counts, and apneas-hypopneas index (AHI).Source: Table 1 · p. 11
• Supplementary Table 2 summarizes the demographics of people with agonal breathing during 9-1-1 calls.Source: Table 2 · p. 11