For patients with acute respiratory distress syndrome (ARDS), mechanical ventilation (MV) is an essential therapy in the intensive care unit. ARDS is diverse condition, the impact of which varies across patients. Every patient has different optimal ventilator settings that may change over the course of treatment, and there is no consensus on how these optimal settings should be found. In particular, the optimal level of positive end expiratory pressure (PEEP) is widely debated. PEEP that is too high or low can cause damage to healthy alveoli, leading to ventilator induced lung injury (VILI). VILI is associated with increased mortality, extended ICU stay, and high cost.

The use of mathematical models to determine patient-specific ventilator settings can reduce the incidence of VILI. There have been many models developed to capture pulmonary mechanics, but they have limitations in lack of ability to capture all relevant physiology, or in complexity and difficulty of implementation. The focus of this research is the development of a model of pulmonary mechanics that does not suffer from many of the disadvantages of previous models.

A nonlinear autoregressive (NARX) model was developed using a complex data set, and contains terms that enable it to fit to all features of the pressure waveform. It captures recruitment and distension across many increasing PEEP steps via an elastance vs. pressure curve that is defined by basis functions. Flow dependent terms allow it to capture viscoelastic effects and fit to an end-inspiratory pause. This model, and slight variations on it were tested on three cohorts of data in total. In many cases the model was compared with the well validated and extensively used first order model (FOM).

Various investigations supported the choice of the NARX model terms. This included using the model for interpolation across a recruitment manoeuvre. The interpolated NARX model fit was consistent across different types of patients, while the FOM performed worse in patients experiencing over-distension at high pressure. Another comparison with the FOM found that the NARX model could more reliably capture expected changes in resistance with PEEP. The NARX model could also identify independent inspiratory and expiratory elastance, due to the flow dependent terms that the FOM does not have.

The NARX model is flexible in its implementation. While it is normally identified in real time using the simple linear least squares method, it was also able to be combined with a modified Gauss – Newton parameter identification method for a spontaneous breathing application. In this case, anomalies in the pressure waveform caused by intermittent patient efforts were able to be removed to enable a more accurate identification of patient parameters.

Aside from patient-specific parameter identification, the main potential clinical use of the NARX model is in predicting the effects of changes in PEEP. An extrapolation of the elastance curve allowed pressure at higher PEEP levels to be predicted. By using partial recruitment manoeuvres as the training data, the NARX model predicted pressure waveforms at higher PEEP levels with significantly lower residuals than the FOM. Since large PEEP changes are not recommended clinically, the most relevant results were the predictions for small PEEP increases of 2 cmH2O. In this scenario the NARX model accuracy was very high.

A statistical classification analysis used the prediction methodology to test the ability of the NARX model to detect when alveolar over-distension is likely to occur with PEEP increases. The analysis considered a pressure threshold above which the risk of over-distension is high. False negatives are potentially much more harmful to patients than false positives, as a false negative means a failure to detect when over-distension will occur with a PEEP increase. Thus, sensitivity was a more important metric than specificity in the analysis. In most scenarios, the NARX model threshold detection had a very high sensitivity and outperformed the FOM, even when compared to a separate method designed to produce the best prediction outcomes from the FOM. However, on one cohort, the parameterisation of the NARX model had to be reduced by reducing the number of basis functions in order to outperform the FOM over large prediction horizons.

An adaptation of the NARX model aimed to capture differences between COPD patients with resultant high auto-PEEP, and non-COPD patients. The adaptation replaced the flow dependent terms with basis functions that enabled linear resistance changes to be captured throughout a recruitment manoeuvre. The model parameters were able to distinguish between the two groups. At low pressure, the high auto-PEEP group had significantly higher modelled resistance, and had elastance curves that indicated a greater proportion of un-recruited lung units. Both of these outcomes were expected due to the airway narrowing and airway closures known to occur in COPD patients.

As in MV, model based glycaemic control can allow personalised care, reduce mortality and improve clinical outcomes. Hence, a side project was undertaken to investigate whether the basis function approaches developed for MV could have potential applications in glycaemic control. The concept was applied to a glucose model to identify a time varying insulin sensitivity (SI) in ICU patients over multiple days. Parameterisation of the model was varied by varying the ratio of basis functions to data points, and this ratio influenced the identified SI profiles that were used to build SI prediction distributions. An analysis determined the appropriate level of parameterisation that resulted in accurate and precise predictions.

The glucose model, the NARX model, and its adaptations all captured clinically relevant patient-specific parameters. The NARX model in particular overcame many of the limitations of previous models, due to the novel use of basis functions to describe elastance, and the use of terms that fit an end-inspiratory relaxation. It achieved this over a range of cohorts that represented a wide variety of patient physiologies and ventilation protocols. The data fitting and prediction outcomes indicate that it has high potential to be useful in diagnosis and disease tracking in a clinical setting.