Automated Cardiac Arrest Detection Using Wrist-Worn Photoplethysmography: External Validation in Patients With Induced Shockable Cardiac Arrest (DETECT-1b)

Survival from unwitnessed out-of-hospital cardiac arrest remains poor due to delayed help.¹ Wearable technology that automatically detects cardiac arrest and activates the emergency medical chain has been proposed as an approach to reduce this delay.^1,2 In the DETECT-1a study, we developed a wrist-derived photoplethysmography (PPG) algorithm for cardiac arrest detection using patient data from induced short‑lasting circulatory arrests, achieving 98% sensitivity.³ In the current DETECT-1b study, we externally validated our PPG-algorithm in patients with induced shockable cardiac arrest, either pulseless ventricular tachycardia (pVT) or ventricular fibrillation (VF).

DETECT-1b was a prospective multicenter study in 50 adult patients undergoing ventricular tachycardia (VT) ablation or subcutaneous implantable cardioverter-defibrillator implantation, while wearing a PPG-wristband (CardioWatch, Corsano Health). During VT ablation, VT was induced using programmed electrical stimulation. Following subcutaneous implantable cardioverter-defibrillator implantation, VF was induced using 50 Hz burst pacing to verify sensing and shock delivery. Invasive blood pressure and ECG were continuously registered as reference. The study was approved by the Medical Ethics Committee East-the Netherlands, and subjects provided informed consent.

Afterward, 2 investigators (R.E./N.T.B.S.) assessed all ECG and blood pressure recordings for ventricular arrhythmias and to determine pulselessness. Induced or spontaneous pVT/VF ≥10 seconds were regarded as cardiac arrest events, and had to be detected by the algorithm. PPG data recorded during the entire procedure were processed by the algorithm developer (K.E.), blinded to cardiac arrest events, using the previously developed cardiac arrest detection algorithm.³ A 10-second detection window was applied to match the brief induced events. Alerts were classified as true positives when associated with pVT/VF; alerts during hemodynamically tolerated VT were considered clinically relevant, and all others were false positives. The data supporting the findings of this study are available from the corresponding author upon reasonable request.

The primary end point was sensitivity for detecting cardiac arrest (pVT/VF), assessed on a per-event basis (all individual events) and on a per-patient basis (first event only). Secondary end points included false positive alarms and the positive predictive value for cardiac arrest detection.

In total, 49 patients were included in the analysis, of whom 7 underwent subcutaneous implantable cardioverter-defibrillator implantation and 43 underwent VT ablation. One patient was excluded due to unavailability of blood pressure data. Median age was 66 years (interquartile range, 51–74), and 41 (84%) were men. Seven (14%) had a history of cardiac arrest, and 34 (69%) of myocardial infarction. Half of the patients had moderately/severely reduced left ventricular function. Twenty-six of the 49 patients had a total of 59 shockable cardiac arrest events: 50 due to pVT and 9 due to VF (Figure). In the remaining 23 patients, no events were induced. In total, 125.5 hours of PPG data were assessed for cardiac arrest alerts.

Figure.

Electrocardiography (ECG), invasive arterial blood pressure, and photoplethysmography (PPG) recordings from 4 patients undergoing subcutaneous implantable cardioverter-defibrillator (S-ICD) implantation or ventricular tachycardia (VT) ablation. A, Ventricular fibrillation (VF). After S-ICD implantation, VF is induced using a 50 Hz alternating current burst. The S-ICD delivers a defibrillatory shock, terminating VF and resulting in return of pulsations. The median duration of all VF events was 21 seconds (interquartile range [IQR], 21–24). B, Pulseless ventricular tachycardia (pVT). Programmed electrical stimulation resulted in pVT. After 12 seconds, an electric shock is delivered, which terminates the VT and results in return of pulsatile blood flow. The median duration of all pVT events was 25 seconds (IQR, 18–36). C, Pulsatile VT. Induced pulsatile VT with a mean arterial pressure of 47 mm Hg and pulse pressure of 14 mm Hg. D, False positive cardiac arrest alert. False positive cardiac arrest alert in a patient undergoing VT ablation. The artifact in the PPG signal followed by short-lasting barely visible PPG peaks likely reflect poor sensor-skin contact. The first normal PPG peak reappears after 18 seconds, and the alert is canceled after 22 seconds. Using 1‑minute PPG signal segments, the algorithm achieved a specificity of 99.9% (95% CI, 99.8%–99.9%) for correctly classifying pulsatile segments as non-cardiac arrest. ECG and arterial blood pressure data were monitored using Sensis (Siemens Healthineers), ICM+ (University of Cambridge), and BARD (Boston Scientific).

In the per-event analysis comprising all 59 events, the sensitivity for detection of cardiac arrest was 92% (54/59; 95% CI, 81%–97%). Sensitivity for VF detection was 100% (9/9; 95% CI, 63%–100%) versus 90% (45/50; 95% CI, 77%–96%) for pVT. In total, 33 alerts occurred in the absence of pVT/VF in 13 VT ablation patients; 24 of these occurred during hemodynamically tolerated (pulsatile) VTs and were considered clinically relevant. Twelve of these episodes occurred in a single patient. The remaining 9 alerts, occurring over 125.5 hours, were classified as false positives (Figure). The positive predictive value for cardiac arrest detection was 86% (54/63; 95% CI, 74%–93%). In the per-patient analysis, considering only the first event per patient (n=26), the sensitivity for detecting VF/pVT was 92% (24/26; 95% CI, 73%–99%). Algorithm performance was not influenced by baseline characteristics.

This is the first external validation in patients of a wearable-based cardiac arrest detection model, demonstrating that wrist-derived PPG reliably detects shockable cardiac arrest, with 100% sensitivity for VF. This aligns with the performance of the model in the DETECT-1a study, in which the model was developed.³ In short, the model continuously monitors the PPG amplitude recorded at the wrist. If the amplitude decreases, the signal quality index is assessed to establish whether PPG peaks are still present. If not, a cardiac arrest alarm is triggered.

The sensitivity we observed is markedly higher than in previous research using data of healthy volunteers and stunt persons with tourniquet-induced pulselessness, but we had higher false positives.⁴ This may relate to the short detection interval used due to the brief duration of induced episodes. Future DETECT studies will focus on assessment of false positives during daily life use and validating performance in nonshockable cardiac arrest and implantable cardioverter-defibrillator patients.⁵

This study included invasive blood pressure as a reference standard for pulselessness and used patient-derived pVT/VF data rather than simulated data, enhancing the generalizability to out-of-hospital cardiac arrest. Limitations include the short duration of arrest episodes and controlled, optimal sensor placement, which may not reflect real-world conditions.

In conclusion, shockable induced cardiac arrest can be accurately detected using wrist-worn PPG with high sensitivity. These findings support further development of wearable-based automated cardiac arrest detection technology, aiming to reduce treatment delays and improve out-of-hospital cardiac arrest survival.

Article Information

Acknowledgments

The authors thank Ruud van Kaam for his assistance with the ECG and arterial blood pressure data collection.

Disclosures

Dr van Royen received a research grant from the Dutch Heart Foundation related to this article. The other authors report no conflicts.

Funding Statement

This research project is financed by public-private partnerships allowance made available by Top Sector Life Science & Health to the Dutch Heart Foundation (Hartstichting) to stimulate public-private partnerships (grant number 2021B006).