责任导师 : Andrea PINNA

助理责任导师 : DRES Martin, MARSALA Christophe

Monitoring regional lung ventilation to predict extubation failure

Mechanical ventilation is a lifesaving treatment for critically ill patients. When the precipitating reason why mechanical ventilation has been resolved, the challenge is to safely and promptly separates the patients from the ventilator. In some patients, the separation from the ventilator is poorly tolerated, a process referred to as extubation failure. Extubation failure occurs in 10% to 25% of the patients and is associated with increased morbidity and increased mortality. To prevent the occurrence of extubation failure, noninvasive respiratory supports are employed. Though, clinicians lack reliable monitoring tools to identify the patient that would most likely benefit from preventive measures.
To address this monitoring shortcoming, we want to use the Electrical Impedance Tomography. This monitoring device uses current injection and voltage measurement of electrodes places around the thorax. From this, images that show the conductivity variation are reconstructed. Applied to the thorax, we can monitor the distribution of the ventilation. This technology offers many advantages, it is non-invasive and can be used continuously at the bedside of the patient. With such technology, we want to monitor patients once they are extubated, and attempt to predict as soon as possible the extubation failure.
For this goal, first, we put in place a clinical study called EXIT, in which patients undergoing extubation were monitored with EIT for 48 hours. During the study that lasted 2 years, 37 patients were included, though 2 were not included in the database due to too much noise in the signal.
During the inclusions time, we set up our Electrical Impedance Tomography (EIT) framework, to pre-process the raw EIT data, reconstruct the images and finally, extracting the EIT features from the images. In total, we use 61 EIT features. Most of them are coming from the scientific literature. But after analyzing the first EXIT patients and discussing with clinicians, we introduce 4 new EIT metrics: lung_area, lung_shape, FlowEIT, RSBIEIT. The last two are based on already existing metrics, used by clinicians that we have adapted to be computed on EIT data. The addition of those features allows to significantly improve the extubation failure’s prediction results of our models (+10.8% of sensitivity and +17.9% of specificity for the failure class). For the learning task, we study 3 different dataset models composed of the EIT data coming from EXIT. The goal of this study is to evaluate the impact of the ventilation variation after extubation on the prediction learning models. Three different inference algorithms are learned on each dataset, they are the Decision tree, the Random Forest and the Support Vector Machine. Our results show that the EIT offers good prediction capabilities. We are not only able to predict the extubation failure, but we are able to predict them hours before the clinician’s re-intubation decision. The sensitivity of the failure class increases throughout the weaning observation. Overall, our failure extubation prediction yields a sensitivity of 0.80 and a specificity of 0.73. This shows the usefulness of the EIT during this crucial period.
The collaboration working on this subject regroup 3 partners with expertise in different fields :

• Laboratoire d’Informatique de Paris 6, LIP6: this is a laboratory part of Sorbonne University which focus on a variety of subjects including such research fields like artificial intelligence and embedded systems. Their contribution to this collaboration is their knowledge of machine learning algorithms, symbolic fusion models, knowledge representation and embedded system design.
• Neurophysiologie respiratoire expérimentale et clinique: this laboratory is also part of Sorbonne University and his located in the hospital of la Pitié Salpétrière. Their focus is on the relationships between the nervous system and the respiratory system. Their medical knowledge will pay a crucial part of the development of the prediction algorithm.
• Bioserenity: the core of this company is on remote healthcare. More precisely, they are developing smart textiles embedding medical sensors that would allow monitoring of patients from home. Therefore, their contribution to this subject is on medical sensors.

2020-2022 刊物

• 2022
  • V. Janiak : “Surveillance de la ventilation pulmonaire régionale pour prédire l’échec de l’extubation”, 博士论文, 答辩 2022-12-6, 责任导师 Pinna, Andrea, 助理责任导师 : Dres, Martin, Marsala, Christophe (2022)
  • V. Janiak, V. Jousselin, G. Tallec, Ch. Marsala, U. Saleem, M. Dres, A. Pinna : “Prediction of the extubation outcome through Electrical Impedance Tomography measurements”, IEEE Biomedical Circuits and Systems (BIOCAS), Tapei, Taiwan, Province of China (2022)
• 2020
  • V. Bonny, V. Janiak, S. Spadaro, A. Pinna, A. Demoule, M. Dres : “Effect of PEEP decremental on respiratory mechanics, gasses exchanges, pulmonary regional ventilation, and hemodynamics in patients with SARS-Cov-2-associated acute respiratory distress syndrome”, Critical Care, vol. 24, pp. 596, (BioMed Central) (2020)