TRADEOFFS IN APPROACHES TO THE VENTILATOR CONTROLLER DESIGN

TRADEOFFS IN APPROACHES TO THE VENTILATOR CONTROLLER
DESIGN
Milos Hauskrecht
Clinical Decision MakingGroup,
MITLaboratory for ComputerScience, Rm421,
545 Technology Square, Cambridge, MA02139
e-mail: milos @medg.lcs.mit.edu
From: AAAI Technical Report SS-94-01. Compilation copyright © 1994, AAAI (www.aaai.org). All rights reserved.
Abstract
This article addresses the issue of knowledgerepresentation
and reasoningtechniques used in the design of a ventilator
controller. It presents two approaches: a model-based
approach that uses domainmodelsand corresponding reasoning techniques, and a goal-oriented approach combining
associative and procedural knowledgein implementing
’compiled’ experience-based response of an expert and
usage of more conventional reasoning techniques. The
approachesare discussed in the context of data provided for
this symposium.
In the last section we will describe the basic
architecture of a controller that is under the developmentin
our lab and that will eventually implementa goal-directed
approach.Its developmentis based on the conjecture that the
goal-directed approachis capable of efficiently solving nontrivial portion of the control problemwithout large increases
in the complexity of the knowledge.
Introduction
The problemof ventilator control has beenat the centre of
attention of AI researches since 70s (VM[1]). However,neither of early worksnor latter worksand/or ongoingresearch
projects have received recognition in the medical community
outside the site they were developed.
The approaches tried in the domainencompassrule-based
approaches( [1], [2]), protocol implementations[3], the use
of domainmodels[4] and their combination. Howeverquestions: like ’wherespecific approachshould be used’ or
’whichapproachis really neededto implementan efficient
controller’ remainto be answered.The focus of the following workis to contribute to the debate and present someof
our views. The discussion will be centered around two
approaches: model-basedthat is rooted in the use of domain
modelsand goal-directed that is anchored mostly on the
experience of the expert.
Ventilator control
The problemof ventilator control is a problemof continous
management
of ventilator settings with regards to the actual
patient state. Thusat everypoint in time the control task is:
’Select the best ventilator settings’.
- a set of possible ventilator moves;
- preference criteria allowing selection of the next moveto
take.
Anexampleof howa clinician deals with the task of ventilator control is illustrated by the sequenceof events that happened during the course of patient monitoring in the AIM
data set:
AIMdata set example(hypoventilation)
while the patient was on the ventilator 2 consecutive blood
gas tests revealed:
1. the decrease in Ph from7.28 at (23:16) to 7.23 (4:30);
2. the increase in PaCO2from 44 (23:16) to 53 (4:30).
This is a clinical picture that indicates hypercarbiat and
worseningacidosis due to decreased alveolar ventilation
(hypoventilation).
The action taken by the clinician was to increase the minute
ventilation by increasing respiratory rate (RR)from 16 to
breaths/minute, probably together with the changein the
peak inspiratory pressure (PIP), resulting in an increase
the tidal volumefrom 72 to 80 ml. Anotherparameter that
could be responsible for the changein tidal volume(but was
not monitored)is the inspiratory - expiratory ratio (I:
ratio).
The values tested after the changein the ventilator settings,
PaCO245 and pH 7.27 (6:11), reflected an increase in the
ventilation.
Control criteria
To construct the controller we wouldlike to ’reproduce’ control activities that clinician performswhens/he is faced with
particular circumstances.In the light of a control task
described abovethis translates immediatelyinto the problem
of acquiring criteria allowingto select the next move.In the
context of the data set it meansknowinghowthe particular
set of ventilator parametervalues wasselected.
Althoughthe objective of the ventilator therapy is well
known(get the patient through the critical disease stage with
the smallest negative effect of the therapy) it is sometimes
Withregards to the control task a control configuration consists of:
- a set of parametervalues describing actual patient and
ventilator state
66
1. Notethat PaCO2
canbe estimatedfrompartial alveolar
CO2pressure that in turn can be approximated
as an endtidal CO2pressure (ETCO2).ETCO2
can be measured
continuoslyand non-invasively.
hard to use these to criteria to decide the beterness of a specific moveor state, e.g. was it better to changeRRto 18 or
20 breaths/minute. Instead the criteria used are often defined
indirectly and morevaguely, e.g. ’try to avoid long ventilation with Fit2 higher than .6 due to possible oxygentoxicity’. Theconsequence
of the fact that criteria are not tight is
that more moves(states) need to be considered equally good,
as there are no preferencecriteria available to differentiate
amongthem. This allows also for greater variation in the
subjective control criteria and can be a reason for someof
the disagreementsin the two expert ventilatory plans.
VCO2
V
D
VT
v
C)~VD
.
V
PaCt2
.(~_
~.J HCO3
Figure 1. Parameter dependenciesin the partial model
for ventilation.
In summary
the control criteria used by clinician are:
- weakand often indirect;
- mostly based on ’experience’;
- use ranges rather than specific values.
- imprecisions inherent in the modeland from parameter
estimates tends to propagate and accumulate in computations.
Model-based approach
Modelsare often used in systems because of their capability
to compactlyapproximatereality. Their main advantage is
that they can serve as a common
reference basis for many
different task, e.g. various inference problemsor explanation.
In the ventilator control domainone can use the physiological modelsof of respiratory mechanics,gas transport in
lungs and in the blood. A piece of such modelreleated to the
hypoventilation exampleis captured by the following formulas and shownin figure 1. (arrows indicate the direction from
ventilator settings to controlled parameters):
[HC03]
pH = 6.1 + log 0.03PC02
In the context of the ventilator controller design the modelbased approach can cover manyactivities one need to use.
Howeverone important feature of the control task - control
criteria cannot be incorporateddirectly into the physiological
model. This is due to the lack of a compactmodelrelating an
individual ventilator moveor patient state to the overall goal
of the therapy. This deficiency can be avoided by constructing a score function for preference ordering on states (see
e.g. [4]). Disadvantagesof such score-function modelare:
- it is hard to construct and it can be very often imprecise;
- secondarycriteria relating states to control movesneed to
be developedand usually require search, thus makingthe
methodinefficient.
Goal-directed approach
A modelcan be used in solving multiple different tasks. On
the other side, the knowledgein the goal-directed approach
represents only features relevant to solve one specific task.
PAC02 = PaC02
1)co2863 = (IAPAC02
(Ta:fT--(z
O
The idea of the goal-directed approach to ventilator managementis to represent and use directly the knowledgeof the
following type:
’howto control ventilator in current situation’,
’whatis the best control step to take next ’.
In other wordsit represents and uses ’compiled’responses of
a clinican to specific situations. A’compiled’control
response can correspond to:
- a changein ventilator settings;
- a sequenceof changes(steps) to execute.
1/= RR. V
T
(/D = RR" VD
where [HC03]stands for the bicarbonate concentration,
PAC02for the partial alveolar C02pressure, ~’co2 for the
C02 production rate, VAfor the alveolar ventilation, Vofor
the deadspaceventilation, I) for the total minuteventilation,
VD for the deadspacevolumeand VT for the tidal volume.
Representingmodelsin the controller one can address tasks
like: ’whatis the minuteventilation neededto reach a
PaCt2of 45’, or ’what is the effect of increasing RRon
PaCt2’.
The knowledgeexpressed in this approach overlaps with the
one discussed in connectionwith control criteria. The major
difference betweenthemis that the goal-directed approach
concentrates on the encodingof one ’optimal’ behavior,
rather than defining the range of ’optimal’ control behaviors.
This is of special advantageas there is no needto encodecriteria rangeswhichare hard to fit to specific values. It is also
morenatural for the clinician to specify and express one
There are also disadvantagesassociated with the usage of
models. Theseare:
- someparameters of the model cannot be measuredand can
be only estimated;
67
behavior.
have started to workon the controller kernel that is oriented
towards implementing goal-directed approach.
Theidea of the goal-directed approachis illustrated in figure
2. Here we assumethat B is an adjustable ventilator parameencoded control
B
’optimal’ range
The controller kernel is the part of the controller that
abstracts from the problemsof input data preprocessing and
action execution. Weassumethat preprocessing phase deals
with the noisy data (signal processing) and that action execution moduledeals with the details of control action execution. Wealso assumethat data are received in sets.
a
I.
II.
IlL ’
Architecture of the controller kernel
Theactivities that a controller mustbe able to performare:
- input data preprocessing
- data interpretation, update state description
- selection of the best action (plan) to take
- executionof the action.
A
Figure 2.’ Optimal’control in the goal directed approach
Thestructure of the controller kernel is shownin figure 3.
ter and A is the only parameterrelevant in determiningits
’optimal’ values. The ’encodedcontrol’ curve reflects the
knowledgeof the goal-directed approach. It defines one specific behaviorthat fits the ’optimal’range for any value of A.
Note that the complexityof an encodingof the curve is
strongly dependent on howvalues of A are broken downto
regions(e.g. I, II, III).
STATE
DESCRIPTION
CONTROL
PLANS
INPUT
Similarly to the association betweensituation (region) and
ventilator parameter value, we can makean association
betweensituation and a sequenceof steps. In this case we
speak about protocols or control plans. A piece of such protocol can look like:
RULES
OUTPUT
Figure 3. Controller kernel structure.
repeat increase peak_inspiratory_pressure(PIP) by
wait 1 minute;
until total_ventilation is approximatelyequal to ventilation_estimate
Major problemsof the goal-directed approachin the ventilator control are related to the problemof breaking downthe
space of all possible states to smaller situations (regions) and
to the need of encodingboundariesfor all such regions.
Most commonproblems are:
- hard to find whereboundaries should be;
- fragmentation (can becometoo large);
- completeness(all possible states must be covered).
State description structure behaves like a memory
that
records the current state of both the controlled system(interpreted) and the controller. Thestate description consist of
numberof variables. One group of variables corresponds
directly to the parametersfrom the input data stream (e.g.
Meanarterial pressure, Respiratory rate). Other variable
groups correspondto: dependentparametersor their estimates that can be computedusing well-defined formulas
(e.g. MAP
during pressure-controlled ventilation, or total
ventilation) or tabular definition, transformedparameters
(e.g. abstraction from quantitative to qualitative values) and
infered parameters.
Conjecture
The goal-directed approachreflects the tradoff betweeneftciency and complexity of knowledgeneeded to encode relevant situation. The conjecture, we are currently pursuing, is
that it is possible to break downthe problemspace along few
dimensionssuch that it will allow to solve at least somenontrivial parts of the ventilator control problemwithoutsignificant increase in the complexityof knowledge.
State description structure is updated in the propagation-like
fashion to ensure the consistencyof the actual state description, i.e. whenthe variable value is changedall dependent
variables are reevaluated automatically. The process of
updating can consist of few propagation sweepsinitiated by
the new data from the input stream or from inferences
throughrules.
For the purpose of testing the viability of the conjecture we
Rulesare used in the process of data interpretation as well as
68
for the purposeof encodingactivities related to control. The
formof rules is obvious. A consequentof a rule consists of
actions that can:
- changethe value of the variable;
- producea control action on the output stream;
- control the executionof control plans.
Actions in the consequentare evaluated in an ’edge-triggered’ fashion (wheneverthe antecedent changesits value to
true), thus reducing the need for reevaluating antecedents on
every cycle.
can be active within the top-level therapy plan from figure 4.
In the future we plan to makea specialization relation
betweensteps and control plans explicit and thus allow for
hierarchical structuring of control plans.
Conclusion
Twopossible approachesto the design of ventilator controller have been evaluated above: a model-basedapproach that
offers robustness and flexibilty for solving several different
reasoningtasks related to the control, and a goal-directed
approach which is experience-based and efficiently generates control responsesfor specific situatios.
Thevital part of the knowledgeused in control tasks deals
with the problemof howto achieve the goal, or howto
behavein specific situations. Such knowledge,whenit consists of moresteps, correspondsto control plans (protocols)
discussed above. Control plans allow for encoding fixed and/
or conditional sequencesof steps in a straightforward way.
Weare currently pursuing the conjecture that the dangerous
increase in the complexitycaused by encoding all relevant
situations in the goal-directed can be avoidedin at least some
non-trivial portions of the ventilator control problem.Totest
this conjecture we have started to workon a controller capable of implementingthis approach.The workis in its initial
stage and there are no results currently available.
Control plans are best represented as transition diagrams
with states and condition/actionpairs attached to transitions.
The meaningof a transition is obvious: a transition is taken
wheneverthe condition is satisfied. Actionson transitions
correspondsto actions in rules. Thetransition diagramof the
top-level plan of the ventialtor therapyis illustarted on figure
4. A state named’increase’ in the figure 4 correspondsto the
Acknowledgements
I wouldlike to thank James Fackler, Isaac Kohaneand Peter
Szolovits for manyhelpful discussions. This work was supported by NIHgrant RO1LM04493.
References
FAILURE
SUCCESS
final states
[1] L.M. Fagan, E.H. Shortliffe, B.G. Buchanan:ComputerBased Medical Decision Making: From MYCINto VM.
In W.J. Clancey, E.H.Shortliffe (eds.), Readingsin Medical Artificial Intelligence, Addison-Wesley,
pp. 241-255,
1984.
start state
EXTUBATION
INCREASE
[2] M. Dojat, L. Brochard, F. Lemaire, A. Harf: A knowledge-basedsystemfor assisted of patients in intensive
care units. Int. Journal of Clinical Monitoringand Computing 9, pp. 239-250, 1992.
WEAN
Figure 4. Top-levelplan of the ventilator therapy.
[3]S. Henderson,et.al.: Performanceof computerizedprotocols for the management
of arterial oxygenationin an
intensive care unit., Int. Journal of Clinical Monitoring
and Computing8, pp. 271-28, 1992
therapy stage, whenthe patient dependenceon the therapy is
increasing, weancorresponds to the opposite process.
Everycontrol plan has one start state and at least one final
state. Currentstate of the active plan is kept in the special
control plan variable that can be explored by other parts of
the system. This allows one to distingiuish the result of the
control plan execution(e.g. successand failure in figure 4),
as well as, to solve tasks like synchronizationof two concurrently active plans.
[4] G.W.Rutledge, et.al.: The design and implementationof
a ventilator management
advisor, Artificial Intelligence
in Medicine5, pp. 67-82, 1993
There is no restriction on the numberof plans that can be
concurrently active. This allows for running two control
plans for mutualyindependentparameters or a control plan
that is in fact a specializationof the state in the other active
plan, e.g. a specific plan to increase the ventilator therapy
69