Logo
ScienceStack
Table of Contents
Fetching ...
Sign In

Needles in Needle Stacks: Meaningful Clinical Information Buried in Noisy Waveform Data

Sujay Nagaraj, Andrew J. Goodwin, Dmytro Lopushanskyy, Danny Eytan, Robert W. Greer, Sebastian D. Goodfellow, Azadeh Assadi, Anand Jayarajan, Anna Goldenberg, Mjaye L. Mazwi

TL;DR

This work tackles the problem of documenting line-access events in critical care by turning a waveform artifact, generated during line-access, into a detectable signal. The authors develop a CNN-based artifact detector trained on fixed-length high-frequency BP waveform windows, deploy it in real-time with sliding windows and Gaussian post-processing, and validate it through static and retrospective evaluations before a prospective deployment. They demonstrate that the artifact carries meaningful clinical information, achieving high performance in static tests and robust operation in streaming contexts, with deployment data showing thousands of detections across hundreds of patients and substantial potential reductions in manual documentation burden. The approach offers a practical, interpretable, and unit-wide scalable solution to improve patient safety and support quality improvement initiatives, while highlighting the value of leveraging noise in physiological signals for actionable clinical insights.

Abstract

Central Venous Lines (C-Lines) and Arterial Lines (A-Lines) are routinely used in the Critical Care Unit (CCU) for blood sampling, medication administration, and high-frequency blood pressure measurement. Judiciously accessing these lines is important, as over-utilization is associated with significant in-hospital morbidity and mortality. Documenting the frequency of line-access is an important step in reducing these adverse outcomes. Unfortunately, the current gold-standard for documentation is manual and subject to error, omission, and bias. The high-frequency blood pressure waveform data from sensors in these lines are often noisy and full of artifacts. Standard approaches in signal processing remove noise artifacts before meaningful analysis. However, from bedside observations, we characterized a distinct artifact that occurs during each instance of C-Line or A-Line use. These artifacts are buried amongst physiological waveform and extraneous noise. We focus on Machine Learning (ML) models that can detect these artifacts from waveform data in real-time - finding needles in needle stacks, in order to automate the documentation of line-access. We built and evaluated ML classifiers running in real-time at a major children's hospital to achieve this goal. We demonstrate the utility of these tools for reducing documentation burden, increasing available information for bedside clinicians, and informing unit-level initiatives to improve patient safety.

Needles in Needle Stacks: Meaningful Clinical Information Buried in Noisy Waveform Data

TL;DR

This work tackles the problem of documenting line-access events in critical care by turning a waveform artifact, generated during line-access, into a detectable signal. The authors develop a CNN-based artifact detector trained on fixed-length high-frequency BP waveform windows, deploy it in real-time with sliding windows and Gaussian post-processing, and validate it through static and retrospective evaluations before a prospective deployment. They demonstrate that the artifact carries meaningful clinical information, achieving high performance in static tests and robust operation in streaming contexts, with deployment data showing thousands of detections across hundreds of patients and substantial potential reductions in manual documentation burden. The approach offers a practical, interpretable, and unit-wide scalable solution to improve patient safety and support quality improvement initiatives, while highlighting the value of leveraging noise in physiological signals for actionable clinical insights.

Abstract

Central Venous Lines (C-Lines) and Arterial Lines (A-Lines) are routinely used in the Critical Care Unit (CCU) for blood sampling, medication administration, and high-frequency blood pressure measurement. Judiciously accessing these lines is important, as over-utilization is associated with significant in-hospital morbidity and mortality. Documenting the frequency of line-access is an important step in reducing these adverse outcomes. Unfortunately, the current gold-standard for documentation is manual and subject to error, omission, and bias. The high-frequency blood pressure waveform data from sensors in these lines are often noisy and full of artifacts. Standard approaches in signal processing remove noise artifacts before meaningful analysis. However, from bedside observations, we characterized a distinct artifact that occurs during each instance of C-Line or A-Line use. These artifacts are buried amongst physiological waveform and extraneous noise. We focus on Machine Learning (ML) models that can detect these artifacts from waveform data in real-time - finding needles in needle stacks, in order to automate the documentation of line-access. We built and evaluated ML classifiers running in real-time at a major children's hospital to achieve this goal. We demonstrate the utility of these tools for reducing documentation burden, increasing available information for bedside clinicians, and informing unit-level initiatives to improve patient safety.
Paper Structure (34 sections, 2 equations, 5 figures, 3 tables)

This paper contains 34 sections, 2 equations, 5 figures, 3 tables.

Table of Contents

  1. Introduction
  2. Related Work
  3. Clinical Context
  4. Noise-generating Process
  5. Signal Processing & Anomaly Detection
  6. Methods
  7. Dataset
  8. Modeling
  9. Streaming Inference
  10. Post-processing Techniques
  11. Evaluation
  12. Static Evaluation
  13. Retrospective Evaluation
  14. Threshold Calibration
  15. Prospective Deployment
  16. ...and 19 more sections

Figures (5)

  • Figure 1: High-frequency blood pressure data from vascular lines in patients provide important information about patient state. Vascular line-access generates artifacts (as shown) - note the size and scale compared to physiological waveform data. We employ a sliding window approach for training a CNN classifier and running real-time inference.
  • Figure 2: Post-processing of predicted probabilities from overlapping sliding-windows are smoothed using a Gaussian convolution and thresholded to generate a distinct prediction for each non-overlapping window. Examples of how individual predictions were compared to ground-truth intervals.
  • Figure 3: Physiological waveform streaming architecture. Real-time data is processed and sent to a message broker. Data for training and model evaluation is stored in a proprietary database. For inference, streaming waveform data is pre-processed and fed to a trained model in a sequential, batched fashion. Model predictions are stored in a SQL database back-end before post-processing and analysis.
  • Figure 4: Total line-access events and Frequency of line-access events (events/hour) from model predictions running real-time between November 2023 and February 2024. We compute each metric at the patient-level and plot them for each bi-weekly period (i.e., each point represents the metric for a single, unique patient). Results are stratified by the inferred acuity-level of each patient, as certain bedspaces are reserved for higher-acuity patients. We observe seasonal trends in line-access that can help inform QI initiatives targeting unit-level line-access rates. We also note a higher-burden of line-access frequency among those with higher-acuity.
  • Figure 5: Mock-up report for CCU Morning Rounds on status of A-Lines and C-Lines at the unit- and patient-level. Patient-level summary statistics reflect which lines each patient has, total duration of each line, access count in the last 24 hours, and time of last access. Values displayed come from real model inferences on real patients from a de-identified date.