A Precedence Scheme for Selection and Explanation of Therapies

ABSTRACT

A general scheme to a1d in the select1on of therapies IS descnbed. A topolog1cal sort1ng procedure with1n a general product1on rule representation IS 1ntroduced The procedure IS used to choose among compet1ng therap1es on the bas1s of precedence rules Th1s approach has a degree of naturalness that lends 1tself to automat1c explanat1on of the cho1ces made A system has been Implemented us1ng th1s approach to develop an expert system for plann1ng ther ap1es for pat1ents d1agnosed as hav1ng ocular herpes s1mplex An abstracted example of the systems output on an actual case IS given ·

I INTRODUCTION

Several knowledge-based programs have been wntten that go beyond d1agnos1s and proceed to recommend treatments. The MYCIN program [ 1] used a production rule scheme to arnve at a d1agnos1s However . it conta1ned spec1al rout1nes and a separate class of rules for the select1on of potent1al therap1es The lack of a general procedure severely limits attempts at automatic explanat1on of therapy The CASNET [2] program linked treatments to the assumed progress1on of the d1sea;;e diagnosed However . a clear explanat1on of the reason1ng beh1nd the cho1ce of therapy was lack1ng Early work on the D1g1talis Therapy Adv1sor [3] gave that program some capability for explanat1on of the chosen therapy It accomplished th1s by translat1ng a trace of the code used More recent work on that subJeCt [4] attempted to prov1de explanat1ons based on knowledge of a consult1ng program generated from an automat1c programmer An extens1ve doma1n model of facts about the applicat1on was requ~red The capabil1ty to ach1eve good performance 1n general has not yet been demonstrated by an automatiC programm1ng approach

Most other expert consultat1on systems attempted to expla1n their cho1ce of treatment 1n terms of a trace of the reasoning process tailored to the applicat1on These approaches do not seem to be easily transferable to other doma1ns Th1s paper describes a procedure that Integrates well w1th a product1on rule scheme and has been found natural for the selection of treatments and the explanat1on of that select1on

·Th1s work IS supported 1n part by the National Institutes of Health through research grants DRR- BRP RR643 and EY 00427

A PRECEDENCE SCHEME FOR SELECTION AND EXPLANATION

OF THERAPIES

John K. Kastner and Sholom M. Weiss

Department of Computer Science Rutgers University New Brunswick, New Jersey 08903

II TREATMENT SELECTION

Often a d1agnos1s will not def1n1tively correspond to a particular treatment Rather . a set of possible treatments will be 1nd1cated The cho1ce of a treatment may 1nvolve 1nformat1on about the d1sease. the pat1ent . the pat1ent"s med1cal h1story. and the treatment 1tself An approach that we have found to be valuable 1n select1on of a therapy is a topolog1cal sort1ng procedure The vanous therap1es be1ng cons1dered are sorted accord1ng to f1nd1ngs and d1agnoses determ1ned for the pat1ent The therapy of cho1ce 1s the f~rst therapy 1n the sorted list

The follow1ng example 1s taken from the model developed for ocular herpes s1mplex 1n collaborat1on w1th Dr Chandler Dawson of the Francis I. Proctor Foundation. Un1vers1ty of Californ1a at San Francisco. The d1agnos1s was reached by us1ng a production rule system based on EXPERT [5) The model 1nd1cated that ant1v1ral drugs were 1nd1cated for treatment The ant1v1ral drugs bemg cons1dered were tnfluorthym1d1ne. v1darab1ne. and 1dOxund1ne The sort procedure was used to choose the appropnate therapy

Define < to be the precedence operator lA < B means that A 1s to be ordered before Bl. In the example. several cntena were used as sort keys to order the drugs

no res1stance < resistancl.'

The above 1nd1cates that res1stance to a drug makes 1t less preferable.

no allergv < mild allergv < moderate allergv < severe allergv

Ev1dence lor history) of a pat1ent s allerg1c react1on to a drug makes that drug less preferable

efflcacvltnfluorthvmldinel <

efl1cacyt v1darahme! <

ef I i cacv11doxund 1 nel

Also the eff,cacy of tnfluorthym1d1ne makes it preferable to v1darab1ne and likew1se v1darab1ne 1s preferable to 1doxund1ne The precedence rules are 1nvoked 1n an ordered fash1on The drugs are to be ordered by r es1stance f~rst . then allergy then eff1cacy For example. the eff1cacy of a drug IS less 1mportant than the allerg1c react1on 1n th1s model The sort procedure is integrated into a rule-based system that performs the d1agnosis and therapy select1on The example that will be used here was abstracted from an actual case The system made the following diagnosis.

D1agnosis

The ev1dence strongly suggests patient severely allerg1c to tnfluorthym1dine. lesion res1stant to 1doxund1ne. corneal ulcer. and active infect1ous herpet1c corneal infect1on

S1nce the pat1ent was diagnosed to be severely allerg1c to vidarabine and resistant to idoxurid1ne the order of preference of the available drugs can be calculated.

vidarabi ne < trifluorthymidine < idoxuridine

Before the treatment recommendat1on can be made. choices must be made concernmg such matters as the route of admm1strat1on. dosage. and length of treatment Here. other precedence sort rules were used 1n conJunction w1th product1on rules to arnve at the complete treatment recommendation.

J_r~atment Recommendation

The ev1dence strongly suggests you should administer vidarabine ointment f1ve times daily. Reexamine 1n 4 to 5 days

Ill EXPLANATION

A maJor benefit of the sort is the ability to readily expla1n the reason1ng beh1nd a cho1ce of therapy. Here. tr ifluorthymidme was chosen because the pat1ent was allergic to vi dar abine and the disease was res1stant to idoxund1ne. By assoc1at1ng explanatory strmgs with each sort key , the followmg explanation of the choice IS automatically output

!r:_l!'!_tment Rationale

V1darab1ne was chosen as the most efficacious ant1v1ral w1th no · ev1dence of allergy or res1stance

In add1t1on. the system IS capable of respond1ng to a restncted set of queries by the user 1n a reasonable fashion

Why vidarabi ne.?

It 1s the most eff1cacious ant1v1ral with no ev1dence of allergy or res1stance

Why not tflfluortlwmidlne?

Ev1dence of severe allergy

Why not ldoxuf!dine.?

V1darab1ne 1s the most eff icacious ant1viral

Alternative treatment.?

The evidence moderately suggests that you should administer trifluorthym1dine drops every two hours during the day. Reexamine in 4 to 5 days.

IV ADVANTAGES AND LIMITATIONS

The sort precedence rules can be easily and succinctly described For the above example. the desired sort 1s descnbed 1n the model by 5 short lines. Th1s replaced an equivalent twenty one EXPERT production rules each of which is composed of mult1ple diSJuncts A b1nary decision tree to accomplish this same task would consist of several thousand nodes

The sort procedure requires a few assumptions of the disease model. The ordenng of importance of the sort keys must be strict For instance. in this case resistance is a more important factor than allergy No Interdependencies between or combinations of the keys is considered. For the application chosen in th1s study. these assumptions have not presented any diff icultieS As part of the development of a general system for therapy planning. this approach 1s be1ng exam1ned for applicability in more complex domains.

V CONCLUDING REMARKS

Th1s paper presents an approach to therapy selection wh1ch lends itself to explanat1on. The precedence scheme can be easily and succ1nctly represented and provides a quahtat1ve approach to therapy select1on that has been found natural for an application to the doma1n of ocular herpes s1mplex The approach has been Integrated 1nto a plann1ng system Although a part1cular domain has prov1ded 1mpetus for the development of the approach presented. the method is general and IS currently be1ng tested in another domain. the treatment of rheumatic diseases

ACKNOWLEDGMENTS

The authors gratefully acknowledge the contributions of Chandler R Dawson. MD. to the development of the model of herpes simplex and Casimir A Kulikowski for h1s careful review of this work