Stochasticity of the respiratory mechanics during mechanical ventilation treatment

Stochastic models have been used to predict dynamic intra-patient respiratory system elastance (Ers) in mechanically ventilated (MV) patients. However, existing Ers stochastic models were developed using small cohorts, potentially showing bias and overestimation during prediction. Thus, there is a need to improve the stochastic model's performance. This research investigates the effect of the kernel density estimator (KDE) parameter tuned with a constant, c on the performance of a 30-min interval Ers stochastic model. Thirteen variations of a stochastic model were developed using varying KDE parameters. Model bias and overestimation were evaluated by the percentage of actual data captured within the 25th - 75th and 5th - 95th percentile lines (Pass50 and Pass90). The optimum range of c was chosen to tune the KDE parameter and minimise the temporal variations of modelpredicted 25th - 75th and 5th - 95th percentile values of Ers (& UDelta;Range50 and & UDelta;Range90) in an independent retrospective clinical cohort of 14 patients. In this cohort, the values of & UDelta;Range50 and & UDelta;Range90 exhibit a converging behaviour, resulting in a cohort-optimised value of c = 0.4. Compared to c = 1.0 (benchmark study model), c = 0.4 significantly reduces model overestimation by up to 25.08% in the 25th - 75th percentile values of Ers. Overall, c = 0.3-1.0 presents as a generalised range of optimum c values, considering the trade-off between data overfitting and model overestimation. Optimisation of the KDE parameter enables more accurate and robust Ers stochastic models in cases of limited training data availability.

Automated summary

Stochastic models for predicting intra-patient respiratory system elastance (Ers) in mechanically ventilated patients have been developed using small cohorts, resulting in potential bias and overestimation. This research investigates the effect of tuning the kernel density estimator (KDE) parameter with a constant, c, on the performance of a 30-min interval Ers stochastic model. By developing variations of the stochastic model using different KDE parameters, model bias and overestimation were evaluated. The optimization of the KDE parameter enables more accurate and robust Ers stochastic models, even with limited training data availability.