Build a neuro-fuzzy continuous protocol representation for ventilator adjustments in acute respiratory distress syndrome patients using fuzzy neural network

Abstract Acute respiratory distress syndrome (ARDS) is on the increase due to many causes such as: corona virus disease 2019 (COVID-19) which worsens the arterial hypoxemia, smoke inhalation, and injuries that cause fluids to collect in the air sacs of the lungs. Standard clinical criteria have been developed to define ARDS severity, while titration protocols guide ventilator settings. The Berlin definition protocol classifies ARDS diagnostic severity based on the $$PaO_2/FiO_2$$ ratio, whereas positive end-expiratory pressure (PEEP) titration is guided by the ARDS Network (ARDSNet) lower and higher PEEP scales. The challenges that however come with physicians manually implementing discrete lookup tables include step-discontinuities at threshold boundaries, performance accuracy, and timely intervention. With the increasing number of patients in the intensive care unit (ICU), physicians may be overwhelmed with the workload, that they may not efficiently control the parameters on the mechanical ventilation. Any slight change or delay in setting the parameters may lead to wrong outcomes. Therefore, the aim of this study was to apply artificial intelligence in guiding physicians during the process of setting ventilator parameters. Fuzzy neural network (FNN) was used in training and conversion of the non-fuzzified ARDSNet PEEP titration tables into a fuzzified continuous inference surface. In this computational study, model performance was evaluated using 5-fold cross-validation and tested on unseen intermediate $$FiO_2$$ setpoints ( $$FiO_2 \\in \\{0.35, 0.45, 0.55, 0.65, 0.75, 0.85, 0.95\\}$$ ), achieving a validation Root Mean Squared Error (RMSE) of 0.018 cm H $$_2$$ O, Mean Absolute Error (MAE) of 0.012 cm H $$_2$$ O, and $$R^2 = 0.9998$$ . The results show that the output of the fuzzified ARDSNet PEEP model is very accurate in comparison with non-fuzzified tables while eliminating abrupt step changes between $$FiO_2$$ increments. Comparative analysis demonstrates that ANFIS provides smooth $$C^1$$ -continuous parameter transitions without the boundary slope discontinuities of linear interpolation or the overshoot oscillations of cubic splines. The potential for reduced clinician workload and error reduction represents a hypothesis for future prospective clinical trials.