A sliding-window based algorithm to determine the presence of chest compressions from acceleration data

Wolfgang J. Kern, Simon Orlob, Birgitt Alpers, Michael Schörghuber, Andreas Böhn, Martin Höller, Jan‐Thorsten Gräsner, Jan WnentData 2022 · DOI

TL;DR

This paper introduces a novel, automated algorithm that accurately identifies chest compressions from accelerometer data, eliminating the need for time-consuming manual annotation. This advancement is crucial for improving CPR quality assessment, enabling large-scale data analysis, and facilitating machine learning applications in emergency medicine.

What Problem Does This Paper Address?

During cardiopulmonary resuscitation (CPR), the effectiveness of chest compressions is a critical factor in patient survival. Traditionally, assessing CPR quality, particularly metrics like Chest Compression Fraction (CCF), requires meticulous manual annotation of chest compression periods from raw data, a process that is both labor-intensive and prone to human error. This bottleneck limits the scalability of CPR quality analysis, especially for large datasets or real-time feedback systems. This paper aims to overcome this challenge by developing an automated algorithm that can reliably detect chest compressions from accelerometer data, making CPR quality assessment more efficient and accessible.

Methods

The core of this paper is a sliding-window based algorithm designed to detect chest compressions from accelerometer data. The process involves several key steps:

  1. Preprocessing: Raw acceleration data is first bandpass-filtered to remove slow baseline drifts, gravitational acceleration contributions, and high-frequency noise.
  2. Sliding Mean: A centered sliding mean is computed over a 1-second interval of the absolute acceleration. This metric tends to be relatively constant and non-zero during compressions and near-zero during pauses.
  3. Derivative Analysis: The algorithm then analyzes the first derivative of this sliding mean. Peaks and troughs in this derivative are identified as potential start and end points of chest compression periods.
  4. Soft Shrinkage: A soft shrinkage step is applied to the derivative to remove minor, non-significant fluctuations.
  5. Candidate Filtering: Potential compression periods are filtered based on three conditions:
  6. The mean absolute acceleration during a pause must be significantly lower (less than 35%) than the average acceleration during the preceding and succeeding compression periods.
  7. Pauses must be at least 1.6 seconds long.
  8. The mean absolute acceleration during the identified compression periods must exceed a certain threshold.
  9. Refinement: For cases with two similar large maxima in the absolute derivative, a weighted mean is used to refine the start/stop marker position, with the weight derived from the absolute derivative over a 1-second interval.

This algorithm is implemented in Python and leverages a web-based interactive plotting tool for manual annotations.

Datasets

The study utilizes data from five exemplary cases obtained from the German Resuscitation Registry (GRR). These cases include continuous-time data such as accelerometry, electrocardiogram, and capnography time series, along with defibrillation event data. The data was collected from ZOLL X-Series defibrillators. Manual annotations of cardiac arrest, return of spontaneous circulation (ROSC)/termination of CPR, and the start/end times of chest compressions were performed by two experienced physicians and reconciled in consensus. The raw data and annotations are made available in a repository.

Key Insights and Ideas

  • Automation is Key: The most significant insight is that reliable detection of chest compressions can be automated using accelerometer data, bypassing the laborious manual annotation process. This has profound implications for the scalability of CPR quality research.
  • Derivative as a Signal: The algorithm's reliance on the derivative of the sliding mean of acceleration is a clever approach to highlight the dynamic changes associated with chest compressions, making them more distinguishable from static or slow-moving signals.
  • Multi-stage Filtering: The multi-stage filtering process, incorporating pause duration, relative acceleration levels, and thresholding, is crucial for achieving high accuracy and robustness against noise and artifacts.
  • Generalizability Potential: While developed for a specific defibrillator model, the authors suggest the underlying principle of using accelerometer data and derivative analysis is likely generalizable to other manufacturers and even thoracic impedance data.

Key Results

The paper presents an algorithm that reliably detects chest compression periods automatically. While specific quantitative performance metrics (e.g., accuracy, precision, recall) are not explicitly detailed in the provided text, the algorithm is described as producing "start and stop markers of single chest compression periods" that can be compared to manual annotations. The accompanying figure (Fig. 1) visually demonstrates the algorithm's workflow and its close agreement with manual annotations, suggesting high fidelity. The primary result is the development of a robust, automated method that significantly reduces the effort required for CPR quality assessment.

Limitations and Future Work

  • Manufacturer Specificity: The authors acknowledge that the algorithm was developed using data from one specific defibrillator manufacturer (ZOLL X-Series). While they expect the general principle to be applicable to others, further validation on data from different manufacturers is needed.
  • Thoracic Impedance: The paper suggests that the algorithm's principles could be extended to analyze thoracic impedance data, which is another common signal recorded by defibrillators. This represents a promising avenue for future research.
  • Broader Application: The authors envision the algorithm's incorporation into general CPR analysis frameworks, facilitating its widespread use in clinical quality management, research registries, and machine learning applications for improving emergency response systems.

Summary updated

Abstract

This publication presents in detail five exemplary cases and the algorithm used in the article (Orlob et al. 2022). Defibrillator records for the five exemplary cases were obtained from the German Resuscitation Registry. They consist of accelerometry, electrocardiogram and capnography time series as well as defibrillation times, energies and impedance when recorded. For these cases, experienced physicians annotated time points of cardiac arrest and return of spontaneous circulation or termination of resuscitation attempts, as well as the beginning and ending of every single chest compression period in consensus, as described in Orlob et al. (2022). Furthermore, an algorithm was developed which reliably detects chest compression periods automatically without the time-consuming process of manual annotation. This algorithm allows for an usage in automatic resuscitation quality assessment, machine learning approaches, and handling of big amounts of data (Orlob et al. 2022).
BibTeX
@inproceedings{kern2022sliding,
  author = {Wolfgang J. Kern and Simon Orlob and Birgitt Alpers and Michael Schörghuber and Andreas Böhn and Martin Höller and Jan‐Thorsten Gräsner and Jan Wnent},
  title = {{A sliding-window based algorithm to determine the presence of chest compressions from acceleration data}},
  booktitle = {Data},
  year = {2022},
  doi = {10.1016/j.dib.2022.107973},
}
Figure