Automated Assessment of Cardiovascular Sufficiency Using Non-Invasive Physiological Data

Xinyu Li, Michael R. Pinsky, Artur Dubrawski

TL;DR

This paper introduces a machine learning approach to automatically assess cardiovascular sufficiency in patients undergoing fluid resuscitation using only non-invasive physiological data. Their novel "Optimized Aggregation of Predictions" framework effectively handles inter-patient variability, enabling accurate predictions even without individual patient baseline data, which is crucial for real-world, out-of-hospital scenarios.

What Problem Does This Paper Address?

Critically ill patients, especially those experiencing hemorrhagic shock, require careful fluid resuscitation. Too little fluid can lead to organ damage, while too much can cause its own set of complications. The key is to accurately determine "cardiovascular sufficiency" – the point where the body has adequate blood flow to meet metabolic demands without being in overt failure. Currently, this assessment relies heavily on invasive monitoring (like arterial pressure and mixed venous oxygen saturation), which are often unavailable in pre-hospital or field settings. This paper aims to bridge this gap by developing an automated system that uses readily available, non-invasive vital signs to assess this crucial physiological state, making it applicable in resource-limited environments.

Methods

The core of the paper lies in two main innovations:

  1. Feature Extraction from Non-Invasive Data: The researchers extract a rich set of features from non-invasive physiological signals. This includes heart rate variability (HRV) metrics from ECG, beat-to-beat geometric features from photoplethysmography (PPG), and derived metrics like Mean Arterial Pressure (MAP), Stroke Volume Variation (SVV), Pulse Pressure Variation (PPV), and Pulse Transit Time (PTT). These features are then aggregated over short time windows (20 seconds).
  2. Optimized Aggregation of Predictions (OAP) Framework: This is the paper's novel contribution to handle inter-patient diversity.
  3. Personalized Normalization: For training, each subject's data is normalized using its own stable "baseline" period. This helps the model learn patterns without being overly influenced by individual physiological differences.
  4. Leveraging Past Baselines: For new, unseen patients (where a stable baseline might not be available), the OAP framework uses the normalization factors (derived from baselines) of previously seen patients. The model makes predictions on the new patient's data, each time normalized using the baseline of a different past patient.
  5. Majority Voting: The binary predictions from these multiple "normalized" views are aggregated using a majority vote. This ensemble-like approach robustly handles the variability inherent in new patients.
  6. Cost-Optimal Threshold Selection: The paper also incorporates a method to select the optimal decision threshold for classifying sufficiency, considering the clinical costs associated with false positives (under-resuscitation) and false negatives (over-resuscitation).

The machine learning model used for classification is a Random Forest, trained using a leave-one-subject-out cross-validation strategy to maximize generalization.

Datasets

The study utilized physiological data collected from a laboratory porcine model of induced hemorrhagic shock and fluid resuscitation. Sixteen healthy female pigs were used. The data included:

  • Beat-to-beat non-invasive estimates of MAP, SVV, and PPV.
  • High-resolution waveform data (250 Hz) from invasive arterial pressure, ECG, PPG, pulse oximetry, and airway pressure.
  • Invasively measured SvO2 and SpO2.

The data was collected across several stages: stabilization, bleeding, waiting, and resuscitation. For model training and evaluation, segments from all stages were used, but performance was specifically evaluated on the resuscitation stage. The authors note that data from 4 subjects was corrupted, leaving 12 subjects for final validation and testing.

Key Insights and Ideas

  • Non-invasive is Sufficient: The study demonstrates that it's possible to achieve clinically relevant accuracy in assessing cardiovascular sufficiency using only non-invasive physiological data, eliminating the need for invasive monitoring in many scenarios.
  • The "Personal Baseline" Problem is Solvable: The biggest hurdle in applying ML to patient data is the vast inter-patient variability. The OAP framework cleverly sidesteps the need for a new patient's baseline by leveraging the baselines of past patients, effectively creating a generalized understanding of physiological states.
  • Ensemble Power for Robustness: Aggregating predictions from multiple slightly different perspectives (each normalized by a different past patient's baseline) acts like an ensemble method, significantly improving robustness and generalization.
  • Clinical Cost Matters: The inclusion of cost-optimal threshold selection highlights that the "best" model performance isn't just about raw accuracy, but about minimizing the most detrimental clinical errors, which can vary depending on the specific care setting.

Key Results

  • The proposed "Optimized Aggregation of Predictions" (OAP) approach achieved an AUROC of 0.929 (±0.021 standard error), which was comparable to a model that did have access to the test subject's personal baseline (AUROC 0.947 ±0.017) and significantly better than a model without any baseline normalization (AUROC 0.892 ±0.019).
  • At a low False Positive Rate (FPR) of 0.030, the OAP approach achieved a True Positive Rate (TPR) of 0.343, meaning it correctly identified 34.3% of sufficient windows while minimizing false alarms. This is close to the best-performing model (TPR 0.387).
  • At a low False Negative Rate (FNR) of 0.023, the OAP approach achieved a True Negative Rate (TNR) of 0.358, correctly identifying 35.8% of insufficient windows. This is substantially higher than the baseline model (TNR 0.195).
  • McNemar's tests showed that the OAP approach performed significantly better than the "Without Personal Baseline" model across various cost scenarios (p-value < 0.001).

Limitations and Future Work

  • Animal Model: The study was conducted on a porcine model, and the authors acknowledge that direct translation to human physiology requires further validation.
  • Small Cohort Size: While leave-one-subject-out cross-validation was used, the relatively small number of subjects (12 for evaluation) is a limitation.
  • Future Work: The authors plan to adapt this work to larger, more complex human patient datasets from intensive care units and emergency medicine settings. They also aim to develop a fully automated, closed-loop resuscitation control system that can dynamically adjust resuscitation modes based on real-time non-invasive vital sign data.

Summary updated

Abstract

For fluid resuscitation of critically ill individuals to be effective, it must be well calibrated in terms of timing and dosages of treatments. In current practice, the cardiovascular sufficiency of patients during fluid resuscitation is determined using primarily invasively measured vital signs, including Arterial Pressure and Mixed Venous Oxygen Saturation (SvO2), which may not be available in outside-of-hospital settings, particularly in the field when treating subjects injured in traffic accidents or wounded in combat where only non-invasive monitoring is available to drive care. In this paper, we propose (1) a Machine Learning (ML) approach to estimate the sufficiency utilizing features extracted from non-invasive vital signs and (2) a novel framework to address the detrimental impact of inter-patient diversity on the ability of ML models to generalize well to unseen subjects. Through comprehensive evaluation on the physiological data collected in laboratory animal experiments, we demonstrate that the proposed approaches can achieve competitive performance on new patients using only non-invasive measurements. These characteristics enable effective monitoring of fluid resuscitation in real-world acute settings with limited monitoring resources and can help facilitate broader adoption of ML in this important subfield of healthcare.
BibTeX
@inproceedings{li2022automated,
  author = {Xinyu Li and Michael R. Pinsky and Artur Dubrawski},
  title = {{Automated Assessment of Cardiovascular Sufficiency Using Non-Invasive Physiological Data}},
  booktitle = {Sensors},
  year = {2022},
  url = {https://openreview.net/forum?id=80h3cVtzL9},
}
Figure