Model-based development of a fuzzy logic advisor for artificially ventilated patients.

This thesis describes the model-based development and validation of an advisor for the
maintenance of artificially ventilated patients in the intensive care unit (ICU). The advisor
employs fuzzy logic to represent an anaesthetist's decision making process when adjusting
ventilator settings to safely maintain a patient's blood-gases and airway pressures within desired
limits. Fuzzy logic was chosen for its ability to process both quantitative and qualitative data.
The advisor estimates the changes in inspired O2 fraction (FI02), peak inspiratory pressure
(PEEP), respiratory rate (RR), tidal volume (VT) and inspiratory time (TIN), based upon
observations of the patient state and the current ventilator settings. The advisor rules only
considered the ventilation of patients on volume control (VC) and pressure regulated volume
control (PRVC) modes.
The fuzzy rules were handcrafted using known physiological relationships and from tacit
knowledge elicited during dialogue with anaesthetists. The resulting rules were validated using a
computer-based model of human respiration during artificial ventilation. This model was able to
simulate a wide range of patho-physiology, and using data collected from ICU it was shown that it
could be matched to real clinical data to predict the patient's response to ventilator changes.
Using the model, five simulated patient scenarios were constructed via discussion with an
anaesthetist. These were used to test the closed-loop performance of the prototype advisor and
successfully highlighted divergent behaviour in the rules. By comparing the closed-loop
responses against those produced by an anaesthetist (using the patient-model), rapid rule refinement
was possible. The modified advisor demonstrated better decision matching than the
prototype rules, when compared against the decisions made by the anaesthetist.
The modified advisor was also tested using data collected from ICU. Direct comparisons were
made between the decisions given by an anaesthetist and those produced by the advisor. Good
decision matching was observed in patients with well behaved physiology but soon ran into
difficulties if a patients state was changing rapidly or if the patient observations contained large
measurement errors.