Simona Turco

Simona Turco

Associate professor at the Biomedical Diagnostics lab, Signal Processing Systems group, Electrical Engineering

Modeling and analysis of patient mechanical ventilation

Perioperative monitoring research

Modelling the lung–ventilator interaction and analyzing ventilation waveforms to estimate respiratory effort and detect patient–ventilator asynchrony.

Overview

Mechanical ventilation stands as a crucial medical intervention utilized to support individuals with compromised respiratory function. Employing a mechanical ventilator, this life-saving technique assists or replaces spontaneous breathing, particularly in intensive care units and emergency scenarios. Despite its indispensable role, mechanical ventilation is not without potential risks and complications. Patients undergoing this intervention may encounter challenges such as ventilator-associated lung injury, infections, hemodynamic instability, and notably, patient-ventilator asynchronies. These asynchronies refer to discrepancies in timing, flow, or volume between the patient’s respiratory efforts and the ventilator’s delivery of breath, posing an additional concern that healthcare professionals must carefully address to optimize patient outcomes.

Modeling of the lung-ventilator system

Accurate modeling of the interaction between the lungs and the mechanical ventilator is of paramount importance for better understanding of the mechanical ventilation process and the more accurate estimation of important quantities such as the patient’s work of breathing. Additionally, the occurrence of patient-ventilator asynchronies (PVAs), leading to patient discomfort, increased work of breathing and also linked to elevated mortality rates, underscores the need for real-time decision support to identify and manage these asynchronies. Motivated by this, we proposed a method for creating a large, realistic, and labeled synthetic dataset. The aim is to facilitate the training and validation of machine learning algorithms capable of detecting a wide range of asynchrony types. The approach is based on a coupled lung-ventilator model: a non-linear lung-airway model is adapted for application in a diverse patient population; additionally, a first-order ventilator model is incorporated to generate labeled pressure, flow, and volume waveforms representative of pressure support ventilation. The model successfully reproduces fundamental measured lung mechanics parameters, and experienced clinicians were unable to differentiate between the simulated waveforms and actual clinical data. This rich annotated dataset is now publicly available at https://doi.org/10.4121/81220350-0f3d-4e0a-86cf-28c904a1cb09.v1.

Electrical equivalents for lungs and ventilator
Electrical equivalents of the ventilator (left) and lung models (right). Adapted from van Diepen et al. J Clin Monit Comput 36, 1739–1752 (2022). View full-size figure

Estimating patient respiratory effort

This model of the lungs and ventilator is also investigated for a better estimation of the patient’s respiratory effort. In fact, although various pressure-based techniques are available for estimating inspiratory effort at the bedside, uncertainty surrounds the accuracy of their estimations. This uncertainty arises because these techniques are largely based on a simplified linear model of the respiratory system, neglecting crucial factors such as the gas compressibility of air and the viscoelasticity and nonlinearities inherent in the respiratory system. We thus performed an in-silico study to compare pressure-based estimation techniques and assess their accuracy using a more sophisticated model of the respiratory system and ventilator. The study evaluated the impact of several parameters on the accuracy of these techniques, including the patient’s respiratory mechanics, positive end-expiratory pressure (PEEP), inspiratory pressure of the ventilator, gas compressibility of air, viscoelasticity of the respiratory system, and the strength of inspiratory effort.

Main finding: In the in-silico comparison, whole-breath occlusion had the lowest average and maximum errors across the patient archetypes evaluated.

On the other hand, techniques based on esophageal pressure demonstrated errors associated with the viscoelastic element in the model, resulting in higher errors compared to occlusion-based methods. The accuracy of esophageal pressure-based techniques was found to be highly dependent on the patient’s pathology and ventilator settings, potentially evolving over time as the patient recovers or deteriorates.

Estimating respiratory system parameters

In more recent work, we investigated an extended equation of motion for estimating respiratory system parameters during mandatory mechanical ventilation. This complements the work on respiratory effort and lung–ventilator modelling described above. Read the published paper by van Diepen et al., Biomedical Signal Processing and Control, 2026.

Automatic detection of patient-ventilator asynchronies

The intricate interplay between patients and mechanical ventilators in clinical settings can sometimes lead to disruptions in the synchronization of their respiratory efforts, giving rise to a phenomenon known as patient-ventilator asynchronies (PVAs). This refers to the mismatches in timing, flow, or volume coordination between a patient’s spontaneous breathing efforts and the assistance provided by the mechanical ventilator. Patient-ventilator asynchronies have attracted increasing attention within the realm of critical care, as their occurrence can have significant implications for patient outcomes. These asynchronies can compromise the effectiveness of ventilation, contribute to patient discomfort, and, in severe cases, may lead to respiratory muscle fatigue or injury. They have also been associated with increased length of ICU stay and mortality. In present clinical practice, PVAs are typically identified and explored through the visual assessment conducted by intensive care physicians. Nevertheless, this method has notable limitations. Firstly, visual inspection is time-consuming and susceptible to subjective interpretation. Secondly, maintaining continuous monitoring through visual means is impractical, potentially leading to the oversight of a significant number of PVAs. Third, accurate detection and classification of PVAs by visual analysis requires trained clinical staff. Here we proposed a machine learning approach for automatic detection and classification of PVAs . The proposed method is based on a U-net architecture, appropriately adjusted for 1-D signals (measured pressure, flow and volume). Based on these input signals, the network learned to recognize the start and end of patient’s respiratory efforts. Comparing these timings with ventilator triggering and cycling enables the detection and classification of different PVAs.

Main finding: In the original evaluation, patient-effort detection achieved sensitivity and precision of 98.6% and 97.3% for inspiratory efforts, and 97.7% and 97.2% for expiratory efforts. Classification accuracy ranged from 90% to 98% across the PVA types evaluated.

U-net architecture
Adapted U-net architecture for automatic detection of PVAs. View full-size figure

Combining simulated and clinical data

To improve the model, we extended the training by using simulated data produced with the model of the lung-ventilator system explained above. We found that the combination of simulated and clinical data for training the algorithm achieved the best performance.

Generating realistic annotated waveforms with VentGAN

A further improvement in detection and classification performance was obtained by refining the simulation through a generative-adversarial network (GAN) approach (VentGAN). Although clinical data with good-quality labels is scarce, large ventilation datasets are available without labels. The idea is to learn the characteristics of clinical data which cannot be modeled by the simulator through adversarial learning. By using the simulated data as input, however, we are able to maintain the known timings, and thus generate a more realistic simulated dataset with labels. The idea is summarized schematically by the figure below. With the exception of ineffective efforts, we found that when using the generated data to train the above-mentioned machine learning algorithm for automatic detection and classification of PVAs, we improve both the detection and classification accuracy of all the investigated PVAs.

GAN approach
GAN framework for generating a large dataset of realistic ventilation data with labels. View full-size figure

Reducing the need for expert annotations

Training asynchrony-detection models requires expert-labelled ventilation waveforms, which are time-consuming to obtain. In a semi-supervised approach, we use distance-based clustering to select waveforms for annotation and project the resulting labels onto similar data.

Main finding: On the evaluated San Matteo dataset, distance-based clustering reduced expert annotation effort by more than 75%, with classification performance close to training with fully labelled data.

Application to a separate Maasstad Hospital dataset also highlighted the importance of annotation consistency. Read van de Kamp et al., IEEE Open Journal of Control Systems, 2026.

PhD theses

Selected publications and resources

  1. van Diepen A, Bakkes THGF, De Bie AJR, Turco S, Bouwman RA, Woerlee PH, Mischi M. A model-based approach to generating annotated pressure support waveforms. Journal of Clinical Monitoring and Computing, 2022. Read the paper.
  2. van Diepen A, Bakkes THGF, De Bie AJR, Turco S, Bouwman RA, Woerlee PH, Mischi M. Evaluation of the accuracy of established patient inspiratory effort estimation methods during mechanical support ventilation. Heliyon, 2023. Read the paper.
  3. van Diepen A, Bakkes THGF, De Bie AJR, …, Turco S. Estimation of respiratory system parameters using an extended equation of motion during mandatory mechanical ventilation. Biomedical Signal Processing and Control, 2026. Read the paper.
  4. Bakkes THGF, Montree RJH, Mischi M, Mojoli F, Turco S. A machine learning method for automatic detection and classification of patient-ventilator asynchrony. IEEE EMBC, 2020. Read the paper.
  5. Bakkes T, van Diepen A, De Bie A, …, Turco S. Automated detection and classification of patient-ventilator asynchrony by means of machine learning and simulated data. Computer Methods and Programs in Biomedicine, 2023. Read the paper.
  6. Hao L, Bakkes THGF, van Diepen A, …, Turco S. An adversarial learning approach to generate pressure support ventilation waveforms for asynchrony detection. Computer Methods and Programs in Biomedicine, 2024. Read the paper.
  7. van de Kamp L, Weller D, Thijssen R, Hunnekens B, Bakkes T, Turco S, den Uil C, Oomen T, van de Wouw N. Reducing annotation effort in patient-ventilator asynchrony detection with distance-based clustering. IEEE Open Journal of Control Systems, 2026. Read the paper.
  8. Sabbah A, Bakkes THGF, De Bie A, Mischi M, Turco S. Multi-task learning for the detection of patient-ventilator asynchrony. EUSIPCO, 2026. Read the paper.

Contributors and partners

Research team

  • Anouk Waeijen-van Diepen (former PhD student)
  • Tom Bakkes (former PhD student, now postdoctoral researcher)
  • Liming Hao (former visiting PhD student)
  • Amin Sabbah (former master student)
  • Nishith Chennakeshava (PhD student)
  • Pierre Woerlee (part-time professor)
  • Massimo Mischi (full professor, head of BM/d)

Collaborators and partners

  • Catharina Ziekenhuis Eindhoven
  • Policlinico S. Matteo (Pavia, Italy)

Funding

  • Eindhoven MedTech Innovation Center (e/MTIC)
  • TKI-HTSM MEDICAID - Top Sector Holland High Tech
  • China Scholarship Council (CSC), National Key Research and Development Project of China
  • Academic Excellence Foundation of BUAA for PhD Students
Partner and funder logos

Back to contents · Perioperative monitoring research · Publications and code