Determining Next Best Action by Knowledge Based AI Agent in Healthcare
This is the last in a series of articles that proposes a method to design intelligent agents in healthcare to computationally answer some of the more complex questions such as which treatment is best for a patient with multiple comorbidity, how do I engage with a member so that he/she renews her policy.
If you have not read the first two articles in the series, you can find them here and here.
In the previous two articles of this series, we have discussed about a formal representation of patient’s state and talked about how we can design Frames in a Thematic Role System in order to create possible future states from an action (in the example we have taken, administration of a medication). We have also seen how this method adopts existing clinical knowledge and data in their current standard.
Let us now see how such a Thematic Role System can enable an intelligent agent to apply appropriate interventions.
Consider, for example, a patient, who had heart issues in the past and who has been prescribed Thiazolidinediones to control Diabetes. The doctor may prescribe diuretics or beta blocker to avoid any cardiovascular complications. Now an intelligent agent would know based on Figure 6 in our last article that one of the complication in the output state can be cardiovascular issue. In that case, the intelligent agent can automatically bring in two Thematic Roles one that administers Diuretics and another that recommends patient monitors his/her weight on regular basis. If we put together the entire picture it looks like the following:
Note that the thematic roles themselves need to be designed in a manner so that the agent can identify the side effect through proper clinical terminology and then search the Thematic role that negates the problem described in the resulting side effect of a state. This can be performed computationally by applying the knowledge contained within the Frames of the Thematic Role System. This state generation does not need to stop after a single step. If the patient has multiple comorbidity, this process can be repeated to generate various possible future states and interventions required to reach the state.
It is also to be noted that once an enterprise defines a standard set of Thematic Roles and their corresponding Frames, it can really use the historical clinical data to build the appropriate state and Thematic Role actions. Such a record can be represented using a simple acyclic, unidirectional graph or a linear tree. If we go back to Figure 1b in the first article, we can now redraw that diagram with slight modifications as the following:
Figure 3b shows the type of graph a healthcare organization can build based on their historical patient records. This provides them a rich set of data that shows individual patient states and the interventions producing those states.
That opens up additional possibilities of using patient centric semantic nets.
Thematic role system enables us to design intelligent agents that can computationally generate multiple states for a patient and can then compare them with other patients who had similar conditions and who went through a similar set of clinical interventions. This allows the intelligent agent to figure out what actions can be taken on a particular patient state in order to have a high chance of recovering from the condition.
If we consider the above Figure 3c, patient A is a reference patient who has gone through a series of interventions described in Thematic Role System a1, a2 and a3 and State A4 is where the patient has fully recovered. An intelligent agent then can find a patient K who was similar to reference patient A. Patient K may need different set of actions based on Thematic Role System resulting in states K3-K8. Now the intelligent agent can use various techniques to determine which of these states are closest to a state in reference patient A. This will help the intelligent agent to optimally choose an action based on thematic role system and ignore other states. The intelligent agent can also select a different reference patient if one of the outcome state is closer to that of a different reference patient.
Similar to the comparison to a positive case where the patient recovered from the condition, the intelligent agent can also compare the state of patient K to that of a patient who did not recover from a condition. In this case the objective of the intelligent agent would be to maximize the distance from the State of patient K and any state of the patient who did not recover.
Depending on the complexity of the use case, the number of computationally generated patient states can be extremely high that can potentially overcome even most generous of computing infrastructure. To alleviate this, using problem reduction techniques in AI, the intelligent agent can prioritize one goal over another and as a result, can target few specific states for further action while discarding others.
Finally, any recommendation of an intelligent agent in this manner will always be explainable since the knowledge-based constructs of Frames and Thematic Roles ensure that every action that is being applied through a Thematic Role System has clear clinical justification and it is derived based on patient’s State. This is perhaps one of the largest advantage of applying knowledge based intelligent agents in the fields of healthcare. Since the intelligent agent uses knowledge constructs to allow usage of existing clinical knowledge, it becomes easy for the agent to explain its actions.
Any intelligent agent in healthcare is expected to aid the human doctor to make better decisions. Once the intelligent agent recommends one or more than one connected set of actions, the human doctor can override it. The intelligent agent in similar fashion can record the individual case and its end outcome. This will also enable the intelligent agent to learn by recording positive and negative cases.
Summary
The techniques that I have described in this set of articles is a general-purpose mechanism to create knowledge-based structure in healthcare in order to build AI agents. The same concepts can be used to solve clinical or non-clinical problems. Once the problem is framed with a knowledge-based construct, there are a number of AI techniques that can be applied formally to arrive at recommendations.
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.
