An expert system for oncology protocol ... - Semantic Scholar

ONCOCIN: AN EXPERT SYSTEM FOR ONCOLOGY PROTOCOL MANAGEMENT

Edward H . S h o r t l i f f e , A. C a r l i s l e S c o t t , Miriam B. B i s c h o f f , A. Bruce C a m p b e l l , W i l l i a m Van M e l l e , C h a r l o t t e D. Jacobs H e u r i s t i c Programming P r o j e c t Departments of M e d i c i n e and Computer S c i e n c e Stanford U n i v e r s i t y , Stanford, C a l i f o r n i a 94305 ABSTRACT oncologists in the treatment of cancer p a t i e n t s . Because t h e o p t i m a l t h e r a p y f o r most c a n c e r s i s not y e t known, c l i n i c a l oncology research i s commonly based on formal experiments that compare the t h e r a p e u t i c b e n e f i t s and s i d e e f f e c t s ( t o x i c i t y ) o f proposed a l t e r n a t i v e d i s e a s e treatments. "Cancer" is a general term f o r many diseases having different prognoses and natural histories. A treatment t h a t is e f f e c t i v e a g a i n s t one tumor may be i n e f f e c t i v e against another. Thus a typical cancer research center may conduct many simultaneous experiments, each concerned w i t h a d i f f e r e n t kind o f cancer and i t s optimal therapy ( i . e . , the treatment plan w i t h the best chance o f cure, remission, or reduction i n tumor s i z e , and t h e l e a s t chance o f s e r i o u s s i d e e f f e c t s ) .

We d e s c r i b e an o n c o l o g y p r o t o c o l management s y s t e m , named ONCOCIN, t h a t is designed to a s s i s t physicians in the treatment of cancer p a t i e n t s . The system is a c t u a l l y a set of programs, one o f which i s a r u l e - b a s e d r e a s o n e r t h a t encompasses t h e necessary knowledge of cancer chemotherapy. Representation and control techniques are d i s c u s s e d , and ONCOCIN i s c o n t r a s t e d w i t h systems that could b e b u i l t u s i n g EMYCIN. Of p a r t i c u l a r interest is the need t o p r o v i d e ONCOCIN w i t h a n interface that w i l l make t h e system acceptable to oncologists. I

Introduction

This report describes an oncology p r o t o c o l management s y s t e m , named ONCOCIN a f t e r i t s domain of e x p e r t i s e (cancer t h e r a p y ) and its historical d e b t t o MYCIN [ 5 ] . The program c o n s i s t s o f a s e t of interrelated subsystems, the principal ones being:

Each of t h e s e experiments is termed a "protocol". P a t i e n t s w i t h tumors of the type being studied, and who are accepted for protocol t r e a t m e n t , are randomly a s s i g n e d t o r e c e i v e one o f two or more p o s s i b l e t r e a t m e n t s . The e x p e r i m e n t requires close monitoring of each patient's clinical response and treatment toxicity. These d a t a a r e t a l l i e d f o r a l l p a t i e n t s t r e a t e d under t h e a l t e r n a t i v e r e g i m e n s , and i n t h i s way the " s t a t e o f - t h e - a r t " i s updated over t i m e .

(1) the Reasoner, a rule-based expert consultant t h a t is the core of t h e system and t h e major s u b j e c t o f t h i s r e p o r t ; and ( 2 ) t h e I n t e r v i e w e r , a n i n t e r f a c e program t h a t c o n t r o l s a high-speed terminal and t h e i n t e r a c t i o n w i t h the physicians using the system.

Each protocol is described in a detailed document, o f t e n 4 0 t o 6 0 pages i n l e n g t h , which s p e c i f i e s the a l t e r n a t i v e t h e r a p i e s b e i n g compared and t h e d a t a t h a t need to be c o l l e c t e d . No s i n g l e p h y s i c i a n i s l i k e l y t o remember t h e d e t a i l s i n even one o f t h e s e p r o t o c o l documents, n o t t o m e n t i o n t h e 30 to 60 p r o t o c o l s t h a t may be used in a major cancer c e n t e r . A l t h o u g h an e f f o r t is made to have the documents available in the oncology c l i n i c s when p a t i e n t s are being t r e a t e d f o r t h e i r tumors, it is o f t e n the case t h a t a busy c l i n i c schedule, c o u p l e d w i t h a complex p r o t o c o l d e s c r i p t i o n , leads a p h y s i c i a n to r e l y on h i s memory when d e c i d i n g drug doses or what laboratory tests to order. Furthermore, solutions for a l l possible treatment problems cannot be s p e l l e d out in protocols. P h y s i c i a n s use t h e i r own Judgment i n t r e a t i n g t h e s e patients, resulting in some variability in treatment from p a t i e n t to p a t i e n t . Thus p a t i e n t s b e i n g t r e a t e d on a p r o t o c o l do n o t always r e c e i v e t h e r a p y i n e x a c t l y t h e manner t h a t t h e e x p e r i m e n t a l design s u g g e s t s , and the data needed f o r formal analysis of treatment results are not always c o m p l e t e l y and a c c u r a t e l y c o l l e c t e d .

Work on ONCOCIN began in m i d - 1 9 7 9 , and t h e system was i n s t a l l e d for prelimianry use i n May 1 9 8 1 . This paper d e s c r i b e s only the Reasoner, a l t h o u g h other system components a r e mentioned when t h e i r i n t e r f a c e w i t h t h e Reasoner i s p e r t i n e n t . We a l s o c o n t r a s t ONCOCIN w i t h EMYCIN [ 8 ] , e x p l a i n i n g why t h e EMYCIN f o r m a l i s m was inadequate f o r our purposes although it did strongly i n f l u e n c e the system's rule-based design. II

Overview of t h e Problem Domain

ONCOCIN is designed to assist clinical -------------------------------------------------T h i s work was supported by t h e N a t i o n a l L i b r a r y o f M e d i c i n e under Program P r o j e c t Grant LM03395 and r e s e a r c h c a r e e r development award LM00048. The program was d e v e l o p e d on t h e SUMEX Computer P r o j e c t at S t a n f o r d U n i v e r s i t y , a shared national resource supported by the Biotechnology Resources Program of t h e NIH under g r a n t RR-00785.

876

The problems we have described reach far beyond the oncology c l i n i c at Stanford Medical Center. There are now several i n s t i t u t i o n s designing protocol management systems to make the d e t a i l s of treatment protocols r e a d i l y available to oncologists and to insure that complete and accurate data are collected . ONCOCIN is s u p e r f i c i a l l y s i m i l a r to some of the developing systems, but i t s research objectives are unique in ways we describe in the f o l l o w i n g s e c t i o n . One o v e r r i d i n g point requires emphasis: in order to achieve i t s goals, ONCOCIN must be used d i r e c t l y by busy clinicians; the implications of this c o n s t r a i n t have pervaded a l l aspects of the system design.

Ill

t r a n s l a t i n g the information i n t o a computer-based format. Knowledge base development has therefore been dependent on the active c o l l a b o r a t i o n of Stanford oncologists. We have started by encoding the knowledge contained in the protocols for treatment # of Hodgkins Disease and the non-Hodgkins lymphomas . In standard use of the system, the physician uses a video display t e r m i n a l , a f t e r examining a patient, to interact with ONCOCIN's dataa c q u i s i t i o n program (the I n t e r v i e w e r ) . The session normally includes: reviewing time-oriented data from the p a t i e n t ' s previous v i s i t s to the c l i n i c , entering information regarding the current v i s i t , and receiving recommendations generated by the "Reasoner" for appropriate therapy and t e s t s . The Reasoner and Interviewer are linked with one another as shown in F i g . 1. In generating i t s recommendation, the Reasoner uses i n i t i a l data about the p a t i e n t ' s diagnosis, data about previous treatment, r e s u l t s of current laboratory t e s t s , plus the p r o t o c o l - s p e c i f i c information in i t s knowledge base. Before terminating an i n t e r a c t i o n , the physician can examine the explanation provided with each recommendation . The physician may approve or modify ONCOCIN's recommendation; any changes are noted by the system and kept available for f u t u r e review. ONCOCIN also provides hard-copy backup to complement the o n - l i n e i n t e r a c t i o n and f a c i l i t a t e communication among c l i n i c personnel.

Design Considerations

The design of the ONCOCIN system has been influenced by the two p a r a l l e l thrusts of the p r o j e c t ' s research aims. The f i r s t is to perform research i n t o the basic science issues of applied a r t i f i c i a l i n t e l l i g e n c e . E f f o r t s are being made to improve the t o o l s c u r r e n t l y a v a i l a b l e , and to develop new t o o l s f o r b u i l d i n g knowledge-based expert systems f o r medical c o n s u l t a t i o n . Other areas of AI research include the extension of methods of interaction between rule-based consultation systems and a large database of timeoriented c l i n i c a l i n f o r m a t i o n , decision making based upon trends over time, and the development of techniques for assessing knowledge base completeness and consistency.

To ensure that busy c l i n i c i a n s w i l l f i n d ONCOCIN fast and easy to use (as well as simple to l e a r n ) , a special terminal i n t e r f a c e has been b u i l t . This Interviewer ( F i g . 1) includes hardware and software features not often found in current medical systems. A customized keyboard incorporates a keypad for most of the entry and command functions. The Interviewer uses a highspeed video display terminal with m u l t i p l e windows; simulating the appearance of the flowsheet that c u r r e n t l y appears in the c l i n i c chart and which is used to record a l l the data required for adequate protocol analysis. As the physician enters flowsheet information, relevant patient data are passed to the Reasoner, which is simultaneously considering the case and preparing a recommendation to pass back to the Interviewer for d i s p l a y . The need f o r a terminal i n t e r f a c e that provides rapid response to user t y p e - i n , while simultaneously satisfying the computational demands of the reasoning process, has led us to design a complex system a r c h i t e c t u r e with asynchronous processes.

The second t h r u s t is to develop a c l i n i c a l l y useful oncology consultation t o o l . One of the primary goals of the project is to demonstrate that a rule-based consultation system with an explanation c a p a b i l i t y can be u s e f u l l y applied in a busy c l i n i c a l environment. While working to achieve t h i s goal, we hope to establish both an e f f e c t i v e r e l a t i o n s h i p with a s p e c i f i c group of physicians, and a s c i e n t i f i c foundation, that w i l l together facilitate future research and implementation of computer-based t o o l s for c l i n i c a l decision making. IV

System Overview

The ONCOCIN system w i l l eventually contain knowledge about most of the protocols in use at the Oncology C l i n i c at Stanford Medical Center. Although protocol knowledge is largely specified in a written document, many questions arise in

We also implemented the complex protocol for t r e a t i n g oat c e l l carcinoma of the lung. Because the oat c e l l protocol is the most complex at Stanford, and it took only a month to encode the relevant rules, we are confident that the representation scheme we have devised w i l l be able to manage, with only minor m o d i f i c a t i o n s , the other protocols we plan to encode in the f u t u r e .

A memo from the MIT Laboratory for Computer Science [7] describes a recent c o l l a b o r a t i o n between MIT and oncologists who have been b u i l d i n g a protocol management system at Boston University [ 3 ] . They are planning to develop a program for designing new chemotherapy protocols. To our knowledge, t h i s is the only other project that proposes to use AI techniques in a clinical oncology system. However, the stated goals of that e f f o r t are d i f f e r e n t from those of ONCOCIN.

We have chosen a representation that w i l l also eventually allow ONCOCIN to offer a j u s t i f i c a t i o n for any intermediary conclusions that the system made in d e r i v i n g the advice.

877

f o r a given parameter

The Reasoner and the I n t e r v i e w e r each run in a s e p a r a t e f o r k under t h e TENEX or TOPS-20 o p e r a t i n g systems, therebv approximating a parallel processing system . Separation of the i n t e r f a c e program f r o m t h e r e a s o n i n g program a l l o w s t h e user t o r e c e i v e prompt a t t e n t i o n when e n t e r i n g f l o w s h e e t values; the Interlisp program, running in background, is not permitted to s l o w down t h e interaction.

;

( 2 ) ONCOCIN does n o t r e q u i r e many c a p a b i l i t i e s p r o v i d e d by EMYCIN.

of

the

( 3 ) Because o f the nature of the i n t e r a c t i o n with the Interviewer, ONCOCIN needs t o operate in a d a t a - d r i v e n mode. Although EMYCIN has l i m i t e d allowance f o r forward chaining of rules, it would be inconvenient to force a largely datad r i v e n reasoning process i n t o t h e EMYCIN format.

B

Representation

Knowledge about the oncology domain is represented using f o u r main types of data s t r u c t u r e : C o n t e x t s , P a r a m e t e r s , R u l e s , and C o n t r o l Blocks. Contexts represent concepts or e n t i t i e s of the domain about which the system needs s t a t i c knowledge, and aid in organizing t h e knowledge base. Individual c o n t e x t s are c l a s s i f i e d by type ( e . g . , d i s e a s e , p r o t o c o l , or chemotherapy) and can be arranged h i e r a r c h i c a l l y . During a c o n s u l t a t i o n , a list of "current" contexts is created as information is gathered. These current contexts together provide a high-level description of the p a t i e n t in terms o f known chemotberapeutic plans. This description serves to focus the system's recommendation p r o c e s s b y s p e c i f y i n g the s i t u a t i o n when a r u l e can be a p p l i e d or when d a t a s h o u l d be requested. Knowledge can be s p e c i f i c to one or more c o n t e x t s a t varying l e v e l s of the contextual hierarchy. Parameters represent the attributes of p a t i e n t s , drugs, t e s t s , e t c . t h a t are relevant for the protocol management t a s k (e.g., white blood c o u n t , recommended d o s e , or whether a p a t i e n t has had p r i o r r a d i o t h e r a p y ) . Each p i e c e o f i n f o r m a t i o n accumulated d u r i n g a c o n s u l t a t i o n is represented as the value of a parameter. There a r e t h r e e s t e p s i n determining the value of a parameter. F i r s t , the system checks t o see i f t h e v a l u e can b e d e t e r m i n e d by d e f i n i t i o n in the c u r r e n t c o n t e x t . If n o t , the " n o r m a l " method of f i n d i n g the value is used: i f t h e parameter c o r r e s p o n d s t o a p i e c e of l a b o r a t o r y data that the user is likely to know, it is requested of the user; otherwise, rules for concluding the parameter are t r i e d . F i n a l l y , the system may have a (possibly context-dependent) d e f a u l t value t h a t i s used i n the event that the

ONCOCIN's Heasoner communicates with the Interviewer during a consultation. The need t o i n t e r a c t w i t h t h i s s p e c i a l i z e d i n t e r f a c e program i s one of several reasons t h a t we chose to b u i l d ONCOCIN f r o m s c r a t c h r a t h e r t h a n t o i m p l e m e n t i t a s a new EMYCIN system [8]. Other important d i f f e r e n c e s between ONCOCIN's a p p l i c a t i o n and t h e domains f o r w h i c h EMYCIN systems have been b u i l t include the f o l l o w i n g : ( 1 ) ONCOCIN requires serial c o n s i d e r a t i o n of each p a t i e n t a t i n t e r v a l s t y p i c a l l y spread o v e r many m o n t h s . Each c l i n i c v i s i t is a new data point, and EMYCIN's d a t a structures do not easily accommodate multiple measurements of the same attribute over time or inference rules based upon assessment o f temporal trends

T h i s same point led to the development o f t h e VM system [ 2 ] , a r u l e - b a s e d program t h a t was i n f l u e n c e d b y EMYCIN b u t d i f f e r e n t i n i t s detailed i m p l e m e n t a t i o n because o f t h e need t o f o l l o w t r e n d s in p a t i e n t s under t r e a t m e n t in an i n t e n s i v e care unit. The development o f s i m i l a r capabilities for ONCOCIN i s a n a c t i v e area o f r e s e a r c h at p r e s e n t . The i n i t i a l p e r s i o n o f t h e s y s t e m , however, encodes o n l y t h e knowledge f r o m t h e p r o t o c o l s and does n o t consider other factors that rely on expert judgment.

Another program, the Interactor, handles i n t e r p r o c e s s communication. There is also a p r o c e s s t h a t p r o v i d e s background u t i l i t y o p e r a t i o n s such a s f i l e backup. The Reasoner i s w r i t t e n in I n t e r l i s p , the I n t e r v i e w e r i n SAIL.

87 8

n o r m a l mechanism f a i l s t o produce a v a l u e , user may be asked to p r o v i d e t h e answer as resort.

ADVISE To make a recommendation a b o u t t r e a t i n g t h e patient: 1) F o r m u l a t e a t h e r a p e u t i c r e g i m e n . 2) D e t e r m i n e t h e t e s t s t o recommend. 3 ) D e t e r m i n e s u g g e s t i o n s about t h e p a t i e n t . 4) Determine the time t i l l t h e p a t i e n t ' s next v i s i t .

or the a last

Rules are the f a m i l i a r productions used i n EMYCIN [ 8 ] and other rule-based systems; t h e y may be invoked in e i t h e r data-driven or goal-directed mode. A r u l e c o n c l u d e s a v a l u e f o r some parameter on the basis of values of other parameters. A rule may be d e s i g n a t e d as p r o v i d i n g a d e f i n i t i o n a l v a l u e or a d e f a u l t value as d e f i n e d above. The r u l e s a r e c a t e g o r i z e d by the c o n t e x t s in which they a p p l y .

To summarize t h e d i f f e r e n c e s between ONCOCIN's r u l e s and t h o s e used in MYCIN and o t h e r EMYCIN systems: (1) Control is separated from domain knowledge, although process i n f o r m a t i o n i s s t i l l codified i n a modular format using control blocks;

A s i n EMYCIN s y s t e m s , r u l e s are r e p r e s e n t e d i n a stylized f o r m a t so t h a t t h e y may be t r a n s l a t e d from L i s p into English for e x p l a n a t i o n purposes . This r e p r e s e n t a t i o n scheme more generally allows t h e system t o " r e a d " and m a n i p u l a t e t h e rules. It has a l s o f a c i l i t a t e d t h e development o f programs t o check f o r c o n s i s t e n c y and c o m p l e t e n e s s o f t h e r u l e s i n t h e knowledge base [ 6 ] . Below is the English translation ONCOCIN r u l e f o r determinjng the value parameter " a t t e n u a t e d dose"

( 2 ) The c o n t e x t u a l i n f o r m a t i o n , which d e f i n e s the setting in which a rule can be a p p l i e d , is s e p a r a t e d f r o m t h e main body of the r u l e and used f o r screening rules when t h e y a r e i n v o k e d ( s e e next s e c t i o n ) ; and

of an f o r the

( 3 ) Rules a r e s u b c l a s s i f i e d t o d i s t i n g u i s h t h e m a j o r mechanisms b y which the values of parameters can be determined ( d e f i n i t i o n a l , n o r m a l , and d e f a u l t r u l e s ) .

RULE075 T o d e t e r m i n e t h e c u r r e n t a t t e n u a t e d dose f o r a l l d r u g s i n MOPP o r f o r a l l d r u g s i n PAVe: If:

Then:

1 ) This i s the s t a r t o f the f i r s t cycle a f t e r a c y c l e was a b o r t e d , and 2) The b l o o d c o u n t s do n o t w a r r a n t dose attenuation Conclude t h a t t h e c u r r e n t a t t e n u a t e d dose is 75 percent of the previous dose.

C

Control

When a user s p e c i f i e s t h e t a s k t h a t ONCOCIN is to perform, the corresponding control block is invoked. This simply causes the steps in the c o n t r o l b l o c k t o b e t a k e n i n sequence. These s t e p s may e n t a i l :

serve as high-level Control blocks d e s c r i p t i o n s o f t h e s y s t e m ' s methods f o r p e r f o r m i n g tasks. Each c o n t a i n s a n o r d e r e d s e t o f s t e p s t o b e used for accomplishing a specific task (e.g., formulating a therapeutic regimen, or c a l c u l a t i n g t h e c o r r e c t dose of a d r u g ) . These c o u l d be viewed as a s c r i p t of t h e events w h i c h o c c u r when a patient is being treated on chemotherapy [4]. Note t h a t t h i s d a t a s t r u c t u r e a l l o w s u s t o s e p a r a t e control descriptions explicitly from d e c i s i o n rules, a distinction that was often unclear in EMYCIN s y s t e m s . The d e s i g n of c o n t r o l b l o c k s was i n f l u e n c e d by the p r o t o t y p e s used in t h e CENTAUR system [1]. The steps in a control block are similar to the p r o t o t y p e ' s c o n t r o l slot. Because we wish to be able t o e x p l a i n any action that ONCOCIN t a k e s , c o n t r o l b l o c k s can be t r a n s l a t e d i n t o English u s i n g t h e same t r a n s l a t i o n mechanism t h a t i s used t o t r a n s l a t e r u l e s . For e x a m p l e :

(1) Fetching data, either by loading previously stored data or by requesting them f r o m t h e u s e r . T h i s causes p a r a m e t e r values to be set, resulting in datadirected invocation of rules that use those p a r a m e t e r s (and t h a t apply in the current context). (2) Determining the value of a parameter. This causes goal-directed invocation of the r u l e s t h a t conclude the value of the parameter (and apply in the current context). Definitional rules are a p p l i e d f i r s t , then t h e normal r u l e s , and if no v a l u e has been f o u n d t h r o u g h t h e s e means, the d e f a u l t r u l e s are tried. If a rule t h a t is invoked in a g o a l - d i r e c t e d fashion uses some p a r a m e t e r whose v a l u e is n o t y e t known, that parameter's value is determined so t h a t the r u l e can b e evaluated. However, c o n c l u d i n g the value o f any p a r a m e t e r , e i t h e r b y t h e a c t i o n of rules o r when i n f o r m a t i o n is entered by the user, may cause data-directed invocation of other r u l e s .

In keeping w i t h the philosophy reflected in o t h e r systems we have d e s i g n e d , ONCOCIN is a b l e to produce natural language explanations for its recommendations. PAVe and MOPP a r e acronyms f o r two of t h e d r u g c o m b i n a t i o n s used t o t r e a t Hodgkins D i s e a s e .

87 9

(3)

I n v o k i n g another

control

(4)

C a l l i n g a s p e c i a l purpose may be d o m a i n - d e p e n d e n t ) .

block. function

(which

The e f f e c t s o f t h i s c o n t r o l mechanism c o n t r a s t w i t h the l a r g e l y backward-chained c o n t r o l used i n MYCIN and o t h e r EMYCIN s y s t e m s . In t h o s e s y s t e m s , a l l rule invocation o c c u r s because t h e v a l u e of a s p e c i f i c parameter i s b e i n g s o u g h t . Rules used t o determine the value of that parameter can be referenced in any order, although ordering is preserved for the assessment of the parameters occurring in the c o n d i t i o n a l statements i n each r u l e ' s premise. Antecedent ( d a t a - d r i v e n ) r u l e s are used when t h e u s e r ' s response to a q u e s t i o n , or ( l e s s commonly) the conclusion from another r u l e , t r i g g e r s one o f t h e s y s t e m ' s f o r w a r d - c h a i n e d r u l e s . These r u l e s can o n l y be used as antecedent r u l e s and typically have single conditions in their premise. I n ONCOCIN (Fig. 2), on the other hand, i n i t i a l control is d e r i v e d from the c o n t r o l block invoked in response to the task selected by the user. F o r w a r d - and back-chaining of rules are i n t e r m i n g l e d , and any r u l e can be used in e i t h e r direction.

o f a f l o w s h e e t s t r u c t u r e f o r d a t a e n t r y has e n a b l e d us to develop a d a t a - d r i v e n system. The e n t r i e s o n the flowsheet define a l l of the important data f o r therapy decisions . Since the doctors use t h e I n t e r v i e w e r t o e n t e r t h e same d a t a t h a t t h e y would normally record anyway, t h e potential frustration of interacting with a goal-directed system is removed.

VI

Why

Artificial

I n t e l l i g e n c e Techniques?

We have l e a r n e d t h r o u g h t h e MYCIN e x p e r i e n c e , and i n b u i l d i n g o t h e r EMYCIN systems a s w e l l , t h a t a major p a r t o f each development e f f o r t has been the encoding of poorly understood knowledge. E n l i s t i n g t h e t i m e and e n t h u s i a s m o f domain e x p e r t s has o f t e n been d i f f i c u l t , yet progress is usually impossible without active collaboration. Thus t h e r e i s g r e a t a p p e a l t o a domain i n w h i c h much o f the needed knowledge is already recorded in t h o r o u g h , a l b e i t l e n g t h y and c o m p l i c a t e d , documents ( v i z . , the protocol descriptions that are w r i t t e n for every cancer therapy clinical experiment). Much o f t h e a p p e a l o f t h e ONCOCIN p r o b l e m domain i s t h e a v a i l a b i l i t y o f d e t a i l e d documents t h a t w e can s t u d y and use f o r knowledge base d e v e l o p m e n t . As we n o t e d e a r l i e r , several other centers have begun t o d e v e l o p p r o t o c o l management s y s t e m s , b u t none has chosen to use t e c h n i q u e s drawn f r o m artificial intelligence. Complicated though the chemotherapy p r o t o c o l s may b e , they are l a r g e l y algorithmic, and o t h e r g r o u p s have been able to encode much o f t h e knowledge using l e s s complex representation techniques. Our reasons for c h o o s i n g a n A I approach f o r e n c o d i n g t h e knowledge o f o n c o l o g y chemotherapy a r e v a r i e d . It should be stressed that all protocols have important loopholes and exceptions; when an a b e r r a n t situation arises for a p a t i e n t being t r e a t e d , the proper management i s typically l e f t unspecified. For example, t h e lymphoma p r o t o c o l s w i t h w h i c h we have been most i n v o l v e d to date include several r u l e s o f the form: IF: THEN:

there i s evidence o f disease extension r e f e r t h e p a t i e n t t o lymphoma c l i n i c

or: IF: THEN:

There a r e t r a d e - o f f s when c h o o s i n g between a system w h i c h i s p r i m a r i l y d a t a - d r i v e n and one w h i c h is g o a l - d i r e c t e d . B a c k - c h a i n e d systems use r u l e s to generate questions, thus facilitating explanations. The d i s a d v a n t a g e , h o w e v e r , i s t h a t a g o a l o r i e n t e d system may appear s l o w t o g e t to the main p o i n t s . On the o t h e r hand, data-driven systems are more d i f f i c u l t to i m p l e m e n t because t h e y must b e a b l e t o h a n d l e e n t r y o f i n f o r m a t i o n i n arbitrary order. In t h e case o f ONCOCIN. t h e use

The broken line in Fig. 2 outlines the p o r t i o n of the ONCOCIN c o n t r o l s t r u c t u r e which i s i d e n t i c a l t o t h a t found i n EMYCIN.

880

there

is significant toxicity to vincristine consider s u b s t i t u t i n g velban

As shown h e r e , the protocols often defer to the opinions of the attending physicians without p r o v i d i n g g u i d e l i n e s o n w h i c h t h e y m i g h t base t h e i r decisions. Hence t h e r e is no s t a n d a r d i z a t i o n of r e s p o n s e s t o u n u s u a l p r o b l e m s , and t h e v a l i d i t y of t h e p r o t o c o l a n a l y s i s i n t h e s e cases i s a c c o r d i n g l y subject to question. Our long-term goal is to d e v e l o p approaches t o t h e s e more complex problems t h a t c h a r a c t e r i z e t h e management o f p a t i e n t s being

The p h y s i c i a n s can i n s e r t new i t e m s i n t o the on-line f l o w s h e e t , as t h e y p r e s e n t l y d o b y hand with the paper f l o w s h e e t . However, these e n t r i e s are not used when generating a therapy recommendation.

treated for cancer. It is when t h e s e issues are addressed t h a t the need f o r A I t e c h n i q u e s w i l l be most e v i d e n t . At t h a t p o i n t the task domain w i l l begin to look s i m i l a r in complexity to the decision problems in a system l i k e MYCIN; r u l e s w i l l have uncertainty associated with them (there is currently n o need f o r certainty weights in the rules in ONCOCIN), and close collaboration with e x p e r t s w i l l once a g a i n b e r e q u i r e d as new r u l e s are written t h a t are not currently recorded in chemotherapy p r o t o c o l s or elsewhere. Even in the s h o r t t e r m , however, A I r e p r e s e n t a t i o n and c o n t r o l t e c h n i q u e s have p e r m i t t e d u s t o keep t h e knowledge base f l e x i b l e and e a s i l y m o d i f i e d . They have a l s o allowed us to develop e x p l a n a t i o n c a p a b i l i t i e s and t o s e p a r a t e k i n d s o f knowledge e x p l i c i t l y i n terms o f t h e i r semantic c a t e g o r i e s .

ACKNOWLEDGMENTS The ONCOCIN System d e s c r i b e d in this report has r e l i e d o n t h e e f f o r t s o f s e v e r a l p h y s i c i a n s and computer s c i e n t i s t s . W e would l i k e t o e x p r e s s o u r g r a t i t u d e t o t h e f o l l o w i n g p r e s e n t and p a s t p r o j e c t members: Paul Chang, J a i Cho, Chuck C l a n t o n , L a r r y Fagan, Phil Gerring, Patricia Pickett, Brandy S i k i c , Motoi Suwa, Randy T e a c h , and C r a i g Tovey. REFERENCES

For t h e p r e s e n t , t h e n , w e a r e t r y i n g to b u i l d a u s e f u l system t o w h i c h complex d e c i s i o n r u l e s can e v e n t u a l l y be added. We know from previous experience t h a t the e n c o d i n g o f knowledge that is not already stated e x p l i c i t l y in protocols w i l l be a r d u o u s and w i l l r e q u i r e a n e n t h u s i a s t i c community of c o l l a b o r a t i n g p h y s i c i a n s . Hence w e r e c o g n i z e the importance of one o f o u r r e s e a r c h goals noted earlier in this report, viz., to establish an effective relationship with a specific group o f physicians so as to f a c i l i t a t e f u t u r e r e s e a r c h and i m p l e m e n t a t i o n o f advanced computer-based c l i n i c a l tools.

VII

1.

A i k i n s , J . S . " P r o t o t y p e s and P r o d u c t i o n R u l e s : A Knowledge Representation f o r Computer Consultations." Doctoral dissertation, Computer Science Department, Stanford U n i v e r s i t y , August 1980.

2.

Fagan, L.M. "VM: R e p r e s e n t i n g Time-Dependent Relations in a Medical S e t t i n g . " Doctoral dissertation, Computer Science Department, S t a n f o r d U n i v e r s i t y , June 1980.

3.

H o r w i t z , J . Thompson, H . , Concannon, T . e t a l . "Computer-assisted p a t i e n t care management i n medical oncology." Proceedings of the 4th Symposium o n Computer A p p l i c a t i o n s in Medical Care, pp. 771-780, W a s h i n g t o n , D . C . , November 1980.

4.

Schank, R. Scripts, Plans, Goals and U n d e r s t a n d i n g , New Y o r k : John W i l e y and Sons, 1977.

5.

Shortliffe, E.H. Computer-Based Medical C o n s u l t a t i o n s : MYCIN, New Y o r k : E l s e v i e r / N o r t h H o l l a n d P u b l i s h i n g Company, 1976.

6.

Suwa, M., S c o t t , A . C . , and S h o r t l i f f e , E.H. "An approach to verifying c o m p l e t e n e s s and consistency in a rule-based expert system". Memo H P P - 0 5 - 8 1 , S t a n f o r d H e u r i s t i c Programming P r o j e c t , S t a n f o r d U n i v e r s i t y , A p r i l , 1981.

7.

Szolovits, P. "Artificial i n t e l l i g e n c e and clinical problem s o l v i n g " . T e c h n i c a l Memo MIT/LCS/TM-140, Laboratory for Computer Science, Massachusetts Institute of T e c h n o l o g y , Cambridge, MA, September 1979.

8.

van M e l l e , W. "A D o m a i n - I n d e p e n d e n t System that Aids in C o n s t r u c t i n g Knowledge-Based Consultation Programs." Doctoral d i s s e r t a t i o n , Computer S c i e n c e Department, Stanford U n i v e r s i t y , June 1980.

Summary

The development of the c o n s u l t a t i o n system addresses basic science questions regarding the optimal representation and u t i l i z a t i o n o f complex medical knowledge. Cancer chemotherapy is a domain that inherently requires consideration of the time course of d i s e a s e , f o r example, and w e a r e o n l y beginning to develop techniques for codifying knowledge that relates to temporal analyses of disease o r o f response t o t h e r a p y . The p r o j e c t seeks t o i d e n t i f y new t e c h n i q u e s for bringing large AI programs to a clinical a u d i e n c e t h a t would be i n t o l e r a n t of systems t h a t are slow or d i f f i c u l t to use. The d e s i g n of an interface that uses both custom hardware and efficient software should heighten the a c c e p t a b i l i t y of the c o n s u l t a t i o n system. ONCOCIN was implemented f o r e x p e r i m e n t a l use w i t h lymphoma p a t i e n t s b e g i n n i n g i n May 1 9 8 1 . Formal e v a l u a t i o n s have been d e s i g n e d t h a t w i l l a l l o w u s to determine b o t h t h e e f f e c t i v e n e s s and t h e a c c e p t a b i l i t y o f t h e system's c l i n i c a l advice. Finally, ONCOCIN is designed to assist physicians with an important clinical domain i n w h i c h t h e r e i s a r e c o g n i z e d need f o r improved d a t a collection and decision making. Furthermore, several other specialty c l i n i c s , at S t a n f o r d and other centers, use time-oriented treatment protocols, and t h e techniques developed here are p o t e n t i a l l y applicable to several c l i n i c a l domains i n a d d i t i o n t o o n c o l o g y chemotherapy p l a n n i n g .

881