Medical Expert Systems

Contexts

Challenges

MYCIN

Consultation system

Explanation System

Rule Acquisition system

Evaluation of Mycin

A Typical Consultation Session

Contexts 

Intelligent  systems, particularly expert  systems for diagnosis and treatment, have been developed for use in a range of medical contexts:

medical practitioners - hospital doctors, nurses, GP’s, consultants, A & E depts, operating theatre, but also nursing home staff, sometimes parents, patients themselves

basic tasks - diagnosis, prognosis, treatment, monitoring

Challenges 

• they may have to decide who gets treated first and who gets sent home - the "triage" situation
• GP's have to decide whether to visit a patient or leave them until surgery
• their knowledge may be out of date
• decision makers in emergency situations may not have access to test results before having to decide on treatment
• they may have fallen into "diagnostic habits"
• they typically have limited time to read journals or go to conferences
• they may have no experience of a particular rare condition and therefore not recognise it or not know how to deal with it
• the specialists whom an a doctor might like to consult may be in different part of country or world
• a treatment may cost so much they hesitate to suggest it
• they may be unsure about their diagnosis and want to check it with someone more experienced
• they may be under time pressure
• they may not be aware of modern drugs, drug interactions, guidelines on drug usage, contra-indications

MYCIN: medical diagnosis using production rules  

• Consultation system
• Explanation System
• Rule Acquisition system

Background

1. is the patient suffering from an infection? (BUT bacteria occur normally; and samples can be contaminated)
2. what’s the organism?
3. which drugs are appropriate? (some drugs too toxic for safe use; no single anitbiotic is generally effective)
4. which one(s) to select?
5. lab tests - quickly for morphology, staining characteristics; later (24-48 hrs) for full identification; later still (1 - 2 days for sensitivity to antimicrobial agents)

• possible type of microbe
• age, weight and sex of patient
• kidney function
• likely effectiveness of antimicrobial agents
• oral, intravenous, intramuscular administration
• dosage

Problems that occur with drug prescription

• non-expert doctors don’t go through this process - prescribe habitual choice
• overprescription (e.g. for viral illnesses) estimated at time at 10-20 times necessary
• wastes money
• can make for resistant strains of organisms

1. Mycin’s Consultation System 

Rule base

• rules stored as premise -action pairs
• premises are conjunctions and disjunctions of conditions
• conditions normally evaluate to True or False, with some certainty factor on the evidence. Rules also have certainty factors. Combined to form new certainty factors
• certainty factor - degree of belief attached to information
• actions are either conclusions (e.g. microbe is type X) or instruction (e.g. remove drug from list of possibles)

Example rule (in doctorese, pseudo-LISP):

IF

1) the stain of the organism is gramneg and

2) the morphology of the organism is rod and

3) the aerobicity of the organism is aerobic

THEN

there is strongly suggestive evidence (0.8) that the class of the organism is enterobacteriaceae

Static and dynamic data structures

Static data structures

These store medical knowledge not suitable for storage as inferential rules: includes lists of organisms, knowledge tables with features of bacteria types, list sof parameters

Parameters = features of patients, bacterial cultures, drugs…

Parameters can be Y/N (e.g. FEBRILE), single value (e.g. IDENTITY - if it’s salmonella it can’t be another organism as well) or multi-value (e.g. INFECT - patient can have more than one infection)

Parameter properties

EXPECT range of possible values

PROMPT English sentence to elicit reponse
LABDATA can be known for certain from Lab data
LOOKAHEAD lists rules mentioning the parameter in their premise (e.g. a rule might need to know whether or not a patient is febrile)
UPDATED-BY lists rules mentioning the parameter in their action (i.e. they may draw a conclusion about the value of the parameter, such as the IDENTITY parameter)

Dynamic data structures store information about the evolving case - the patient details, possible diagnoses, rules consulted:

Example piece of dynamic data:

To  evaluate the premise of the rule mentioned above:

($AND (SAME ORGANISM-1 GRAM GRAMNEG)

(SAME ORGANISM-1 MORPH ROD)

(SAME ORGANISM-1 AIR AEROBIC))

from following data

ORGANISM-1

GRAM = (GRAMNEG 1.0)

MORPH = (ROD 0.8) (COCCUS 0.2)

AIR = (AEROBIC 0.6) (FACUL 0.4)

Total value of whole expression is lowest certainty value = 0.6

Conclusion is therefore:

Control structure

MYCIN first attempts to create a "patient context" containing information about the case, then tries to compile a list of therapies for the context.

It uses a backward chaining mechanism, reasoning back friom the goals it want sto prove to the data it has, rather than vice versa. The overall goal is "compile a list of therapies".

Questions can be prompted by the invocation of rules, to find out necessary data, to avoid unnecessary questions.

2. The Explanation System 

MYCIN can answer questions about HOW a conclusion was reached and WHY a question was asked, either after a consultation or while it is going on.

It does this by manipulating its record of the rules it invoked, the goal it was trying to achieve, the information it was trying to discover.

Can also answer general questions (e.g. what would you prescribe for organism X?) by consulting its static data structures.

3. The Rule Acquisition System 

Experts can enter new rules or edit existing rules. The system automatically adds the new rule to the LOOKAHEAD list for all parameters mentioned in its premise, and to the UPDATED-BY list of all paramters mentioned in its action.

Evaluation of MYCIN 

Initial evaluation: five experts agreed with 72% of its recommendations over 15 cases

In later experiment, MYCIN’s performance was rated against senior and on-senior doctors:
 

A Typical Consultation Session 
 

1. Patient Information

1) Patient’s name

PT538

2) Age

34 years

3) Sex

Male
 

Diagnosis

6) Please enter results of pending cultures in table:

SITE CULTURE# DATE EXAMINED

CSF 1234 30.1 YES

10) Do you suspect PT538 may have an infection at a site from which you have not obtained specimens?

No

15) Has PT538 recently had objective evidence of abnormal neurologic signs (e.g. seizures, coma) documented by physician?

Yes

The CSF culture will be considered to be associated with meningitis.

Cause

17) Does PT538 have an abnormal chest x-ray?

No

18) Has PT538 been exposed to any contagious diseases recently?

No

27) Please enter CSF findings in table:

WBC %PMN Protein CSFglucose Periph-glucose

100 99 280 5 UNKNOWN

35) Is meningitis a hospital-acquired infection?

Yes

39) is PT538 a burns patient?

Yes

I have found evidence (RULE545 RULE557)that treatment should cover for e-coli causing meningitis

I have found evidence (RULE545 RULE557 RULE578)that treatment should cover for pseudomaonas aeruginosa causing meningitis

I have found evidence (RULE545 RULE557 RULE578)that treatment should cover for Klebsiella pneumoniae causing meningitis

I have found evidence (RULE507)that treatment should cover for Diplococcus-pneumoniae causing meningitis
 

Treatment

It is important to cover for the following infection(s) and associated organism(s):

Meningitis

<item 1>e-coli

<item 2>pseudomaonas aeruginosa

<item 3>Klebsiella pneumoniae

<item 4>Diplococcus-pneumoniae

41) Does PT538 have a clinically significant allergic reaction to any antimicrobial agent?

No

43) PT538’s weight in kilograms:

70

Recommendation 1

My preferred therapy is as follows:

In order to cover for items <1 2 3 4>

give the following in combination:

1) AMPICILLIN

Dose 3.5g q4h IV

2) GENTAMYCIN

Dose 119 mg q8h IV

Since high concentration of penicilllins can inactivate aminoglycosides do not mix these two antibiotics in the same IV bottle.