THE ADAPTIVE CONTROL 
OF THERAPEUTIC PROCEDURES 
N. P. Thompson," B. Widrow" and C. Schade" 
If one wishes to optimally control any system which 
exhibits a nonexplicit time variable delay and a nonex- 
plicit time variable response (gain), he must consider 
adaptive control. 
The control approach being used models the animal's 
response by means of a tapped adaptive transversal 
filter. Each tap is separated from the next by a delay. 
The output at each tap is weighted and summed with 
all of the other tap outputs. Thus when there is a 
change in the drug rate being administered to the ani- 
mal, an equivalent scaled voltage analog change is in- 
troduced into the filter. As the response appears in the 
animal's blood pressure, it is compared with the 
summed output of the filter's taps. If there is disagree- 
ment in the two outputs, the filter's weights are adjusted 
(adaption) to bring the two outputs into better agree- 
ment. After several passes of the filter, its response will 
be a faithful model of the animal's response. When an 
error develops between the blood pressure desired and 
the actual blood pressure, the "inverse" of the model's 
transfer function is taken to determine what should 
occur with regards to the vasopressor's administration 
to achieve the desired change in blood pressure. This 
change is then implemented. The animal's response to 
this change is automatically employed to up date the 
model if necessary. 
This procedure can be applied to situations having 
more than one input and response. In such cases, it 
may optimize one response while holding the others 
within acceptable bounds by causing appropriate 
changes in the therapeutic agents being administered. 
INTRODUCTION 
There are several ways one might approach 
the automatic control of therapeutic agents. 
First, the operator may use the controller as a 
simple switch, closing the loop and carrying out 
all of the analyses himself. Thus he may decide 
that a given drug should be administered for a 
set period at a specified rate. The controller 
merely carries out these orders. Such a process 
hardly deserves being called a control system. It 
* Palo Alto Medical Research Foundation Palo Alto, California. 
** Stanford University, Stanford, California. 
has the advantage that the controller is not free 
to do much ; so the faint of heart and distrust- 
ing might find such an approach appealing. Of 
course, such a controller is also of little help. 
The next higher level of controller that one 
might choose closes the loop using some simple 
unchanging algorithm. This algorithm might 
consist of some heuristic plan. For example, if 
A less than B, give 25 cc of C, or it might be a 
deterministic function such as seen in ordinary 
control systems. Here the result (output) is 
compared with the desired result and an error 
value determined. This error is added to or sub- 
tracted from the input to the system under con- 
trol thus bringing the output back to the 
desired level. Such control systems have the ad- 
vantage that one need only ask for a specified 
response and the controller will hold that re- 
sponse, if possible. The problem with such sim- 
ple controllers is that in order to assure stabil- 
ity they must be very slow (as slow as the worst 
case). Therefore, they cannot hope to be opti- 
mum except in the worst case. This objection 
also holds for the simple approach mentioned 
first where the operator closes the loop himself 
if he doesn't wait long enough before making a 
control decision, instability will result. In addi- 
tion, such approaches as these must assume that 
the biological response will be maximal to as- 
sure stability. So also on this second basis such 
simple control systems cannot be optimum ex- 
cept in the best case. Thus such simple control 
systems can never be optimum because in the 
worst case the gain will not be optimum and in 
the best case the delay will not be optimum. 
To summarize, one cannot have optimum and 
stable control in the face of a nonexplicit time 
variable delay and/or gain unless he is pre- 
pared to alter his control strategy to fit changes 
that occur in the gain and delay of the system 
1207 
