At the forefront of Artificial Intelligence

Home » Articles » Natural Language Processing » Applications/Code

MARS: A Virtual Doctor

By T. Harrision, J. Matthews, J. Greensmith, B. Wootton and A. Scott

This page is primarily here to explain the MARS (Medical Automated Response System) parser. All user documentation, database code and the illness database can be found in the zip file at the bottom of this page.

The MARS Parser

The MARS parser works by taking a sentence, removing all extraneous information (upper cases, punctuation etc.) and then creating a sentence tree from the words. The root of the tree is always a node that contains the complete sentence. Nodes can contain words or can denote phrases such as a noun phrase (MEWT_NP). The details of noun phrases and others can be found in the documentation. Let us look at a debug print-out for "My leg really hurts.":

Address:         0x007A8080
sWord:           my leg really hurts 
iType:           MEWT_S (2048)
iModifier:       0
child:           0x007A9DE0
next:            0x00000000
parent:          0x00000000

    Address:         0x007A9DE0
    sWord:           - n/a -
    iType:           MEWT_NP (2)
    iModifier:       0
    child:           0x007A8030
    next:            0x007A9F50
    parent:          0x007A8080
    
        Address:         0x007A8030
        sWord:           my
        iType:           MEWT_DET (16)
        iModifier:       0
        child:           0x00000000
        next:            0x007A9FD0
        parent:          0x007A9DE0
        
        Address:         0x007A9FD0
        sWord:           leg
        iType:           MEWT_N MEWT_DN (128)
        iModifier:       19
        child:           0x00000000
        next:            0x00000000
        parent:          0x007A9DE0
        
    Address:         0x007A9F50
    sWord:           really
    iType:           MEWT_ADV (64)
    iModifier:       1
    child:           0x00000000
    next:            0x007A9EC0
    parent:          0x007A8080
    
    Address:         0x007A9EC0
    sWord:           hurts
    iType:           MEWT_V (256)
    iModifier:       2
    child:           0x00000000
    next:            0x00000000
    parent:          0x007A8080

Now take a look at the query that is generated from this input:

Address:         0x007A9DA0
sSymptom:        leg 
uLocation:       ML_LEG 
uDuration:       MD_ONEDAY
uSeverity:       MS_SEVERITY4
next             0x00000000

Note how the location has been parsed and how "really" has modified "hurts" to produce a higher severity (MS_SEVERITY4 = 3). Let us briefly look at another example of the query system - this query structure was generated from "My left wrist is fractured.":

Address:         0x007A9CE0
sSymptom:        wrist
uLocation:       ML_HAND ML_LEFT
uDuration:       MD_ONEDAY
uSeverity:       MS_SEVERITY4
next             0x00000000

Notice how the location allows for position based information (left). Also note that the word "fractured" was taken to be an indicator of severity not as a symptom. This is perhaps not the best case for all instances but it works in this scenario.

Conclusion

While MARS is not the best diagnostic program, the code is good to learn from. Due to time constraints, the team couldn't implement all the symptom and location checks we had wanted but it does do its job relatively well (spots symptoms and other information, finds the closest match in its database and returns a result).

Again, all the binaries, documentation and source code are available in the zip file below.

- James Matthews (2/10/01)

mars.zip (132Kb) - All binaries, source and documentation.

Submitted: 15/05/2001

Article Toolbar

BibTeX entry