Human control system interaction

Mark,

It depends on which type of interaction you are considering. Do you mean simply controlling the unit, or dealing with the hmi on a more "emotional" basis?

Consider what your control objectives are and what sort of input data you are expecting to handle.

What sort of interaction? Do you mean needing to have the machine adjust for the patient's activity level, or the patient "fighting" the machine, or what?

For that matter, what's your background? Are you a medical doctor looking at control loops for the first time, or a control engineer looking at lungs?

I assume that you've done a thorough search in the bio-tech literature?

Adaptive-fuzzy-neural network control has archived almost cult status, meaning that newbies want to apply it everywhere, even if the problem is one of the many ones where a garden variety PID controller, applied correctly and well tuned, is the absolute best that you can ever expect.

I tend to view AFNN controllers with a fairly jaundiced eye, but I will say that if you can get markedly better performance out of your ventilator when there's a technician standing there and twiddling the knobs vs. when it's just the machine then you might have an application for fuzzy control.

Is this for a "Respirator"?

Browse the web-sites of companies like Penlon and InterMed.

May not be the best technology. I know someone else has already suggested you need to consider the measurable parameters and what they indicate, then know what you are able to control and the effect that control has on the measured parameters. You will need to formalise your map of that a bit then maybe a clearer path will show itself.

One technique that has been applied to particularly knotty problems in electrical and electronics engineering is "Signal Flow Analysis". This builds a map of the flows of signals/information in a system and can, like any good map, indicate the shortest route to a solution.

The only reference I have on the subject is:-

Tiitle: Signal Flow Analysis Author: J.R. Abrahams and G.P. Coverley Publisher: Pergamon Press (1965) ISBN: not stated LoCCC No: 65-16375

You might want to look at passivity based control. Neville Hogan at MIT developed one passivity based technique called "Impedence Control" specifically for control of devices that interact with humans. For a more theoretical and broad description of passivity based control look for the work of Chris Byrnes out of Washington University.

Yet another possible approach is feedback domination which is a very interesting combination of passivity based control with some differential geometric concepts. That's out of Wei Lin's group at Case Western Reserve University and he published an article about two or three years ago in IEEE Trans. Aut. Cont. demonstrating it's application to underactuated and nonsmooth systems.

Good luck,

Rich

Hi Michael,

The work I have done shows that the machine has to be triggered by the patient during spontanious breathing. So the machine is always in lag of the intent. But if the patient coughs or the leak to the mask changes (i.e. with posture) then false triggering occurs.

Even when the triggering is working properly if the support pressure is too high it changes the patients interaction with the machine causing him to fight the ventilator.

Hence there is a whole range of problems with the hardware lagging the intent breath by breath and then more subtle problems of whether the support is at the right level over time. I can control the machine fine but need some formal method (state machine?) of deciding what is happening from the data and taking approperate action

Thanks for your help,

Mark.

Hi Tim,

I'm with you on the Adaptive-fuzzy-neural network bla bla hype and also find PID to be the best solution most of the time which how I am controlling the machine now.

I'm a chartered electrical engineer working on an engineering/medical Phd and with 12 years practical experience in embedded systems/control theory the engineering side is easy. It's the human bit that is proving difficult.

I've got a PID control providing pressure support to patients using a mask. The problem is that if the pressure is to high, delivered too fast or the timing is out the patient significantly changes the way in which they interact with the machine. On top of that some people have a higher tolerance to using this type of machine and will try and sync with it rather than fight it so there is a great varity of responses to the same machine. This leads to a conservative controller to avoid instability which in turn increases the controller lag, decreasing the tolerability for the patient!

Still smiling from your end note about the best control coming from a clinian twiddling knobs being the best option. It is!! but I cant capture what he does with the limited sensing options available.

My thoughts at the moment are to write a huge state machine that passes down setpoints to the PID controller. Using some sort of temperal logic to try and spot events in the data and then take some appropreiate action. It will work but feels like a bodge to me. I'm sure this sort of thing comes up in other industrys, they must have a formal solution (surely..)

Thanks again,

Mark.

Certainly if there were a way to measure intent (elastic around the patient's chest? Neural sensors on the patient's neck?) that would be a cool thing.

Your arguments _do_ point to some kind of adaptive and/or fuzzy control. Your patients will present a highly variable set of plant behaviors, which calls for adaptive control, your detection of incidents like coughing or mask slipping may be much easier to describe linguistically rather than mathematically.

I would certainly look at an adaptive controller that tries to make some conclusions about the patient's behavior and adjust the controller based on that. If this lets you tighten up the PID tuning it would be a good thing. Even this may have to include fuzzy elements -- you'd certainly like to detect when the patient is really panicking and dump the parameters back to some ultra-conservative setting.

You can always console yourself that you _are_ supposed to be contributing some new and useful element to human knowledge; if it were easy someone else would have worked this topic as a master's thesis years ago.

I would consider getting an ambubag (bag ventilation mask), and attempting to ventilate someone for a few minutes. After about 10 minutes with a marginal patient, you should have a pretty good idea of how to pick out when they are going to breathe, and even some idea on how to tell how much to give. You might be able to look at the mask pressure to determine this. Keep a nominal pressure (a cm or two) on the mask. You can filter out coughs, and also add a latency period after coughs, then start detection again.

I'm not sure whether PID or fuzzy sets are the best here. I would consider using a profile based control based upon the breathing cycle.

Michael

Hi Tim,

Spot on! Bands round the waist have been tried (too much interferance from posture changes) and there are many people trying to develop a neural sensor to detect respiratory intent, but nothing so far. Think I'll spend a month looking at Fuzzy logic although I'm not impressed by what I know about the field so far.

The most frustrating thing is that there are about a dozen commercial manufacturers of ventilators and the doctors using them dont talk about one being better than another despite the wide range of design implementations and software. Their attitude is that the patient is either supported or not which makes me wonder if any incremental change will be noticed...

Only 5 months till I finish and then I'll get a proper job again. Cant wait to return to the real world :-)

Thanks again,

Mark.

--snip--

Have you tried asking the respiratory technicians, or the patients? The information may not do much good without buy-in from the doctors, but certainly the _patients_ would have some opinions!