<p>The aim of this thesis was to test the capability of Artificial Neural Networks (ANN) to estimate respiratory mechanics during mechanical ventilation (MV). ANNs are universal function approximators and can extract information from complex signals. </p><p>We evaluated, in an animal model of acute lung injury, whether ANN can assess respiratory system resistance (R_RS) and compliance (C_RS) using the tracings of pressure at airways opening (P_AW), inspiratory flow (V’) and tidal volume, during an end-inspiratory hold maneuver (EIHM). We concluded that ANN can estimate C_RS and R_RS during an EIHM. We also concluded that the use of tracings obtained by non-biological models in the learning process has the potential of substituting biological recordings.</p><p>We investigated whether ANN can extract C_RS using tracings of P_AW and V’, without any intervention of an inspiratory hold maneuver during continuous MV. We concluded that C_RS can be estimated by ANN during volume control MV, without the need to stop inspiratory flow.</p><p>We tested whether ANN, fed by inspiratory P_AWand V’, are able to measure static total positive end-expiratory pressure (PEEP_tot,stat) during ongoing MV. In an animal model we generated dynamic pulmonary hyperinflation by shortening expiratory time. Different levels of external PEEP (PEEP_APP) were applied. Results showed that ANN can estimate PEEP_tot,stat reliably, without any influence from the level of PEEP_APP.</p><p>We finally compared the robustness of ANN and multi-linear fitting (MLF) methods in extracting C_RS when facing signals corrupted by perturbations. We observed that during the application of random noise, ANN and MLF maintain a stable performance, although in these conditions MLF may show better results. ANN have more stable performance and yield a more robust estimation of C_RS than MLF in conditions of transient sensor disconnection.</p><p>We consider ANN to be an interesting technique for the assessment of respiratory mechanics.</p>

The aim of this thesis was to test the capability of Artificial Neural Networks (ANN) to estimate respiratory mechanics during mechanical ventilation (MV). ANNs are universal function approximators and can extract information from complex signals. We evaluated, in an animal model of acute lung injury, whether ANN can assess respiratory system resistance (RRS) and compliance (CRS) using the tracings of pressure at airways opening (PAW), inspiratory flow (V’) and tidal volume, during an end-inspiratory hold maneuver (EIHM). We concluded that ANN can estimate CRS and RRS during an EIHM. We also concluded that the use of tracings obtained by non-biological models in the learning process has the potential of substituting biological recordings. We investigated whether ANN can extract CRS using tracings of PAW and V’, without any intervention of an inspiratory hold maneuver during continuous MV. We concluded that CRS can be estimated by ANN during volume control MV, without the need to stop inspiratory flow. We tested whether ANN, fed by inspiratory PAW and V’, are able to measure static total positive end-expiratory pressure (PEEPtot,stat) during ongoing MV. In an animal model we generated dynamic pulmonary hyperinflation by shortening expiratory time. Different levels of external PEEP (PEEPAPP) were applied. Results showed that ANN can estimate PEEPtot,stat reliably, without any influence from the level of PEEPAPP. We finally compared the robustness of ANN and multi-linear fitting (MLF) methods in extracting CRS when facing signals corrupted by perturbations. We observed that during the application of random noise, ANN and MLF maintain a stable performance, although in these conditions MLF may show better results. ANN have more stable performance and yield a more robust estimation of CRS than MLF in conditions of transient sensor disconnection. We consider ANN to be an interesting technique for the assessment of respiratory mechanics.