Model-based neuro-fuzzy control of FiO2 for intensive care mechanical ventilation

cc1073 CCJ Meeting abstract Model-based neuro-fuzzy control of FiO2 for intensive care mechanical ventilation Kwok HF Mills GH Mahfouf M Linkens DA

Department of Surgical & Anaesthetic Sciences, Sheffield University, and the Intensive Care Unit, Royal Hallamshire Hospital, Sheffield S10 2JF, UK

Department of Automatic Control and Systems Engineering, Sheffield University, Sheffield S1 3JD, UK

Critical Care 21st International Symposium on Intensive Care and Emergency Medicine Meeting abstracts 21st International Symposium on Intensive Care and Emergency Medicine Brussels, Belgium

20–23 March 2001

1364-8535 2001 5 Suppl 1 P002 10.1186/cc1073 15 1 2001 2 3 2001 cc-5-s1-p002

The knowledge-based approach to fuzzy logic control of mechanical ventilation on the ICU can be prone to bias in the experts' knowledge and errors resulting from poor communication during rule-base derivation. Therefore, a different approach was explored in the development of a fuzzy controller to control the inspired oxygen fraction (FiO₂). The performance of such a controller was compared with the performance of the clinicians.

Method

(1) The development of a neuro-fuzzy controller

This was developed by training a neural network to generate an optimal change in the FiO₂ in order to achieve a target arterial oxygen tension (PaO₂) on a mathematical model of the gas exchange system (SOPAVent). The neural network learnt the relationship between the blood gases, FiO₂ and PEEP and other ventilator settings. This was done by exposing the neural network to the blood gas results produced by applying a range of FiO₂ and PEEP values to the SOPAVent model. This first neural network was then combined with another neural network which represented a fuzzy logic rule-base. The fuzzy rule-base consists of a set of 'If ..., Then ...' statements based around combinations of FiO₂, PEEP and PaO₂. The fuzzy rule-base was then adjusted by changing the weights of the neuro-controller (which correspond to the 'Then ...' part of the fuzzy rules) during neural network training. The neuro-controller output is equivalent to the output from a fuzzy inference system of three inputs (the difference between the actual PaO₂ and the target, the PEEP level and the FiO₂).

(2) Comparing neuro-fuzzy and clinicians' control

The scenarios were based on the data from three real patients with sepsis in the ICU. Seventy-one blood gases, ventilatory settings and respiratory parameters at the sampling times were presented to nine consultant intensivists. They were asked to optimise the PaO₂ of the patient scenarios in the simulator by adjusting the FiO₂. Similarly, the neuro-fuzzy controller was presented with the same data and asked to adjust the FiO₂. The impact of these changes on the patient's PaO₂ was then calculated using the SOPAVent model. The FiO₂ adjustments and corresponding new PaO₂ levels were compared to see how close were the decisions of the clinicians and the neuro-fuzzy controller.

Results

These are shown in Table 1. The differences were not statistically significant.

Table 1

FiO₂ (%)

PaO₂ (kPa)

Mean

Median

Mean

Median

Clinicians

44.60 ± 11.63

45.00

14.62 ± 4.08

13.78

Neuro-fuzzy

43.95 ± 11.03

42.20

14.12 ± 2.69

14.43

controller

Conclusion

The control of PaO₂ provided by the neuro-fuzzy controller was similar to the clinicians' control. Neural networks can provide an alternative means for fuzzy rule-base derivation and tuning for ventilator control.

This project was funded by EPSRC Grant no. R/M96483.