Imparting Knowledge and Generating Predictions
Knowledge Based AI Agent in Healthcare
In the first article of this series, we have established a patient centric semantic net and discussed structure and possible content of such a net. In this article, we are going to discuss frames and thematic role systems and why they are important in the functioning of an intelligent agent.
Frames and Thematic Role Systems
‘Frames’ are essential representation of common knowledge. They are typically top-down knowledge-based structures that sets expectation about one or more events. For example, consider the Diabetes type 2 drug Thiazolidinediones. While it reduces blood glucose level by increasing insulin sensitivity, it has potential side effects such as
weight gain, heart failure and heart attack2. As a result this medication should not be prescribed to patients with kidney disease or heart problems. This knowledge can be constructed in a frame that looks like the following:
Today, all of these knowledge are encoded in various drug databases that are available commercially. For the purpose of this discussion we are representing this rule as a Frame. However this knowledge can also be imparted in the form a rule. As we shall discover soon, the source or mode of delivery of the knowledge either as a rule or as a frame is irrelevant to the topic in context.
Next, we talk about Thematic Role Systems. Thematic Role Systems are essentially frames that represent an action and sets specific context and expectation around the action. For example, let us consider the action of medication administration. We would expect additional implicit information about the action, e.g. what is the patient’s condition, which medication, how it is being administered, is there any expected side effect etc. A thematic role system that represents the above information is provided below. As one can see it provides a generalized representation of the action of drug administration without being specific to a drug. The frame structure has been simplified on purpose. The frame can be as rich or as simple as the use case requires.
In the above frame, you can see the various ‘aspects’ of the drug administration are empty as it is a generic representation. If one considers the act of taking oral Thiazolidinediones, the above frame may look like the following after populating relevant information.
Generating Expectations
If we combine the frame representation of Thiazolidinediones in Figure 2a with Thematic Role System of administration of Thiazolidinediones, we can generate a set of expectations (predictions) about the future state of the patient.
Now we go back to the generalized state transition diagram that we created as part of article 1 and apply the expectations of administration of Thiazolidinediones to generate expected patient states:
What this shows is that there are four patient states expected based on the action of Thiazolidinediones oral medication administration.
An intelligent agent can assign probabilities of each of these outcomes. These probabilities can be derived from existing clinical studies or can be derived from existing patient records, if available. In addition, as the patient progresses through states 1–2, 1–3, 1–4 the intelligent agent may be able to adjust the chances of various future states based on the information gained from the states that has already been realized.
There is a key difference between adopting a thematic role-based approach and the way systems are designed currently. The thematic roles allow a healthcare organization to classify current and future actions into specific categories. Such actions can be clinical procedures and non-clinical interventions. Further, clinical procedures can be invasive and non-invasive. The Thematic Role System allows a structured representation of various clinical/non-clinical actions and interventions, in the context of a use case, and generate future states coming out of these actions without writing explicit query and/or rules.
Similarly for a healthcare marketing use case, the Thematic Role System can allow an organization to define actions such as ‘phone outreach’ or ‘email outreach’ or ‘customer service messaging’ and generate expectations of a consumer’s state after each of these actions.
As we’ll see in the next part of the article, this allows such actions to be chained as a sequence of events and examine various possible outcomes at every stage without explicitly writing logic to determine all intermediate states. Needless to say writing such logic explicitly factoring in the vast number of disease conditions and clinical/non-clinical interventions would be very expensive and would be of very high maintenance. Instead, this approach enables intelligent agent to select and apply relevant actions and examine the patient’s future conditions.
Ontology for Filling Up Frames
Before I wrap up this article, I would like to talk about the how one can go about defining the actions for Thematic Roles and creating the Frames for Thematic Roles. Frames and Thematic Role Systems need to be filled in a manner so that they can be applied on a Patient States to generate next set of probable Patient States. In order to computationally generate such states, it is critical that the actions as well as the codes that go in the Frames of the Thematic Roles are standardized. Domain ontology plays a critical role in ensuring a standardized set of code sets and nomenclature is applied.
There are a number of Ontology projects that has shown potential for broader adoption in healthcare. For example, Systematized Nomenclature of Medicine Clinical Term Ontology (SCTO) shows clinical concepts at specific level of abstraction. One can extract the relevant actions and create appropriate Frames for Thematic Role Systems. Once the action is decided, most of the content of the Frame can be filled in from current standard coding standards and drug databases.
A detailed discussion on source data to fill in the Frame is beyond the scope of the article, however I am happy to point you towards right direction below1.
Summary
In this second part of this article, we have seen how one can create Frames in a Thematic Role System keeping specific clinical actions in mind in order to generate expected patient states when an action is taken. Such actions lead to computationally generate future patient states, and thus enable application of additional Thematic Role System Frames as a chain of events and generate various possible future patient states.
The key points are
1) A Thematic Role System provides an efficient framework for generating patient state by applying clinical actions
2) Most of the rules needed are already well documented in various clinical literatures and drug databases and are currently in use in various organizations. Such rules can be extracted and injected or configured in an intelligent agent.
3) Usage of standardized clinical terminology and standards is critical for efficient usage of the methodology. Various ontology projects, already completed, can provide a good foundation for defining the actions of the Frames under Thematic Role System.
In the next section of the article, we are going to talk about how the states generated by applying Thematic Roles can be chained and how decisions can be taken based on comparison of past learning data set.
Any viewpoint or opinion expressed in this blog is strictly personal and belong to the blog author. Unless it is explicitly stated, Such viewpoints or opinions may not be construed to be representative of any organization that the author is associated with currently or was associated with in the past in either personal or professional capacity.
