Qualitative modelling and learning in KARDIO

(File Last Modified Wed, May 29, 2002.)

• Review: Qualitative modelling and learning in KARDIO
• Background
• Details
• KARDIO knowledge-acquisition cycle
• Compression aiming at improved comprehensibility
• Hierarchical model
• Use ML to construct the qualitative models
• Contribution
• Comments

Review: Qualitative modelling and learning in KARDIO

Background

• KARDIO is an expert system for electrocardiographic(ECG) diagnosis of cardiac arrhythmias. The problem is stated as: given a symbolic, qualitative description of the patient's ECG, find the discorders in the patient's heart that could have caused this ECG.
• Qualitative modelling in association with the 'knowledge-acquisition cycle' in KARDIO is applicable as a knowledge acquisition paradigm to many other problem areas.

Details

KARDIO knowledge-acquisition cycle

• Knowledge from the experts and literature
• Model of the heart: Deep knowledge, intensional representation. (25KB)
• Arrhythmia ECG base: Surface, extensional representation. (5MB) ----- deduction: by exhaustive simulation of the medel.
• Compressed rules: Surface, comprehensible. (25KB) ------ induction from the arrhythmia ECG base
• Feedback: compressed rules can be compared with the original domain knowledge, which offers a powerful possibility for knowledge validation

Compression aiming at improved comprehensibility

• Inductive learning program is used to summarize the examples generated from the arrhythmia-ECG base.
• Learning was only used as a representation compression technique in which no generalisation actually occurred.
• Compressed rules are well understandable and correspond to those in the standard medical literature.

Hierarchical model

the model of the heart can be viewed as a relation: *model( Disorders, Manifestations)*

Necessity: although the model(stated in logic) is in principle possible to execute in both direction, the combinatorial complexity of running it in the inverse direction (generate-and-test) is much higher than in the forward, simulation model. One way to circumvent this difficulty is to improve the search performance by introducing a hierarchy into the model.

Hierarchical model has the form: *model(Level, Disorders, Manifestations)* where Level is an integer 1,2,... denotes the abstraction level.

Abstraction is obtained by applying some transformation rules, namely:

• Replace a set of (possibly interconnected) components of the detailed model by a single component.
• Omit a variable in the detailed model.
• Replace the value of a variable at the detailed level by a more abstract(coarser) value; this entails a hierarchy of domain values of variables.

the hierarchy is exploited as follows:

• To solve the diagnostic(manifestations-->disorders) task at some level L, the given manifestations are abstracted to level L-1.
• Then the simpler diagnostic task is solved at level L-1 (further abstractions if exist can be used).
• The resulting solution at this level is then used to guide the search at level L.

This hierarchical formalism can be applied as a generally applicable technique of logic programming to reduce combinatorial search in problems that render themselves to useful abstraction.

Use ML to construct the qualitative models

An integrated system is used for machine-aided construction of the essential part of the KARDIO model from examples (this is the compressed model of the original hierarchical KARDIO model which provides the compressed, understandable representation). The system consists of:

• A propositional type learning program as a basic induction machine.
• A logic debugging algorithm
• An interpreter for qualitative models

The learning of components progresses through two stages:

• Initial stage : initial hypotheses about the components are induced from initial examples; these hypotheses, the structure and background predicates constitute the current _model hypothesis_ .
• Debugging stage : the current model hypothesis is executed on example inputs to the model. The hypothesis is then updated whenever this execution produces results different from those expected by the teacher.

Contribution

KARDIO is a demonstration of the use of deep knowledge in the form of a qualitative model in building a complex knowledge base.

The basic exercise in qualitative modelling for ECG diagnosis was extended in several directions:

• extraction of compact and comprehensible rules
• hierarchical modelling
• machine learning supported construction of qualitative models.

Comments

• The knowledge acquisition cycle disscussed in the KARDIO system inevitably reminds me about the approach we proposed in the MBRE01 paper: the knowledge farming--mining--validation cycle to achieve better decision making.( Note -difference: different assessment, non-exhaustive simulation, ``learnt theories'' is constraint on domain theory)
• The point is: learning in the KARDIO was primarily as a compression rather than generaliation in which no new knowledge was derived. The summarized knowledge was the same as that used to construct the model although it has a computer-based representation, which provided a way for comparison and validation.
• In our case, we didn't construct a model but reuse a model (COCOMO), knowledge obtained by summarization(leaning) can actually change the model's behavior while feeding back. Nevertheless, there is no essential difference between the two cases: since the knowledge base is generated by the model, nothing new can be invented. Does that mean: besides the validation/comparison value, knowledge summarized or rules induced from the model is trivial and obvious? Can't think of any defence for that right now...

Note : ml4re paper(first 2 sections)-- the defence of the above.