Development and evaluation of an artificial intelligence‑based decision support system to optimise mechanical ventilation in acute respiratory failure

Page 512 of 858 Intensive Care Medicine Experimental 2024, 12(Suppl 1):87 Reference(s) 1. Gros, Tania Ojuel, et al. "Intoxicaciones agudas graves en UCI: rasgos epidemiológicos, clínicos y estándares de tratamiento." Atalaya Médica Turolense 11 (2017): 28–35 2. Juárez-Aragón, Guillermo, et al. "Características clínicas y epidemiológicas de intoxicaciones graves en una población adulta que ingresa a una unidad de cuidados intensivos." Gac Méd Méx 135.6 (1999): 669–675. 3. Palazón Sánchez C, Segura Pérez J, Renedo Villaroya A, Palazón Sánchez EL, Pardo Talavera JC, Felices Abad F. Intoxicaciones agudas graves en un servicio de medicina intensiva durante doce años. Rev. Esp. Salud Pública [Internet]. 2000 Feb [citado 2024 Abr 07]; 74( 1): 00–00. Disponible en: http:// scielo. isciii. es/ scielo. php? script= sci_ artte xt& pid= S113557272 00000 01000 07& lng= es. Topic: Poisoning/Toxicology/Pharmacology 001001 Development andevaluation ofanartificial intelligence‑based decision support system tooptimise mechanical ventilation inacute respiratory failure J. Wittenstein1, M. Scharffenberg1, F. Fischer1, R. Theilen1, J. Li2, T. Kramer1, T. Koch1, S. Vahdati2, M. Gama De Abreu3, R. Huhle1 1Department of Anesthesiology and Intensive Care Medicine, University Hospital Carl Gustav Carus Dresden, Dresden, Germany; 2InfAI, Institute for Applied Informatics e.V., Dresden, Germany; 3Department of Intensive Care and Resuscitation, Cleveland Clinic Main Campus, Cleveland, United States of America Correspondence: J. Wittenstein Introduction: Finding an individualised mechanical ventilation (MV) strategy to ensure gas exchange while avoiding ventilator-induced organ damage is challenging in acute respiratory distress syndrome (ARDS). Objectives: We aimed to develop a reinforcement learning-based decision support system (IntelliLung DSS) to dynamically suggest optimised ventilator settings using historic patient data and evaluate its performance in a porcine ARDS model. Methods: Following approval of the institutional review board at the Medical Faculty Dresden (BO-EK-423082021), data from invasively ventilated adult patients with ARDS or pneumonia admitted to the intensive care units (ICU) of the University Hospital Dresden were used to develop the IntelliLung DSS. We randomly selected 80% of the data for training and 20% for the clinical evaluation. We used a discrete batch-constrained deep Q-learning (BCQ) approach to suggest optimised settings for positive end-expiratory pressure (PEEP), inspiratory oxygen fraction (FIO2) and respiratory rate (RR). The patient state was described using 12 variables of MV, gas exchange, haemodynamics and lab results as multidimensional time series in 10-min time steps. The rewards function of the algorithm was defined as ICU mortality and short-term pulmonary and physiological rewards within safe ranges. For prospective performance evaluation four pigs were anaesthetised after approval (AZ: 25-5131/522/24) and sequential lung injury was induced using surfactant depletion (LAVAGE), injurious MV (VILI) and continuous intravenous lipopolysaccharide infusion (LPS). A fixed tidal volume of 6ml/kg was used throughout the experiment. PEEP, FIO2 and RR were set in each lung injury block in randomised order according to either the IntelliLung DSS or ARDS network protocol (ARDSnet, low PEEP table for PaO2/FIO2 > 175mmHg and high PEEP table for PaO2/FIO2 < 175mmHg at the respective baseline)[1] for 180min each (cross-over design). Respiratory and haemodynamic variables were recorded every 15 min. Statistical analysis was performed with SPSS using a general linear model for repeated measurements (mean ± standard deviation). Results: Data from 942 patients (653 male, 289 female; 64 ± 16years, ICU mortality 25%, year of admission: 2010–2020) were used for algorithm development and clinical evaluation. In the clinical evaluation, the IntelliLung DSS more frequently suggested lower FIO2 (≤ 40%) and PEEP (≤ 10 cmH2O) and more frequently RR ≤ 18 /min compared to the clinician. In the animal study, RR was not significantly different between IntelliLung DSS (30 ± 2 min−1) and ARDSnet (31 ± 4 min−1, P = 0.757) (Fig.1). In the LAVAGE block, FIO2 according to IntelliLung DSS was higher than during ARDSnet (43 ± 22 vs 37 ± 21%, P = 0.038), whereas in VILI and LPS blocks there was no significant difference between IntelliLung DSS and ARDSnet strategy (Fig. 1). In the VILI block, PEEP according to IntelliLung DSS was higher compared to ARDSnet (7.4 ± 1.3 vs 5.2 ± 0.7 cmH2O, P < 0.001), whereas in LAVAGE and LPS block there was no significant difference between IntelliLung DSS and ARDSnet (Fig.1). In the LAVAGE block, the ratio of peripheral oxygen saturation (SpO2) to FIO2 was significantly lower during IntelliLung DSS than during the ARDSnet strategy, but not in the VILI and LPS block (Fig.2). There was no significant difference in mechanical power (Fig.2), mean arterial pressure (Fig.2), cardiac output and arterial pH between IntelliLung DSS and ARDSnet strategy regardless of the lung injury block. Conclusions: The IntelliLung DSS was applicable both clinically and experimentally. In the retrospective clinical evaluation, IntelliLung DSS proposed a ventilation strategy with lower FIO2 and PEEP values compared to the clinician. In the animal study, IntelliLung DSS allowed the individualisation of MV similar to the ARDSnet settings, avoiding extreme PEEP values. Fig. 1 (abstract 0001001) Mean and standard deviation. IntelliLung DSS, mechanical ventilation according to the IntelliLung decision support system; ARDSnet, mechanical ventilation according to the ARDS network protocol; PEEP, positive end-expiratory pressure; FIO2, inspiratory oxygen fraction; LAV, lung injury induced by surfactant depletion; VILI, lung injury induced by injurious mechanical ventilation; LPS, lung injury induced by intravenous lipopolysaccharide infusion; ns, not significant Fig. 2 (abstract 0001001) Mean and standard deviation. IntelliLung DSS, mechanical ventilation according to the IntelliLung decision support system; ARDSnet, mechanical ventilation according to the ARDS network protocol; SpO2, peripheral oxygen saturation; FIO2, inspiratory oxygen fraction; LAV, lung injury induced by surfactant depletion; VILI, lung injury induced by injurious mechanical ventilation; LPS, lung injury induced by intravenous lipopolysaccharide infusion; ns, not significant; *P < 0.05 Reference(s) 1. 1. Brower RG, Lanken PN, MacIntyre N, et al. (2004) Higher versus lower positive end-expiratory pressures in patients with the acute respiratory distress syndrome. The New England journal of medicine 351:327–36. https:// doi. org/ 10. 1056/ NEJMo a0321 93 2. This work was supported by an institutional grant of the EKFZ FOR DIGITAL HEALTH at the Medical Faculty Carl Gustav Carus Dresden, Germany