TL;DR
This paper introduces a novel "bias towards relevance" for plan recognition in complex domains like trauma management, where multiple, shifting goals are common. Instead of minimizing the number of goals, it prioritizes goals that are contextually relevant to the current situation, enabling more accurate interpretation of physician actions and better clinical decision support.
What Problem Does This Paper Address?
Plan recognition, the task of inferring an agent's intentions from their actions, typically relies on an inductive bias to prune the vast space of possible explanations. A common bias is to minimize the number of goals used to explain observed actions. However, in domains like multiple trauma management, patients have numerous, interacting, and evolving goals (e.g., diagnosing a head injury while simultaneously managing internal bleeding). In such scenarios, a single-goal minimization bias is inappropriate and can lead to misinterpretations. This paper addresses the challenge of developing a plan recognition system that can effectively handle these complex, multi-goal environments, ensuring that the inferred plans are clinically meaningful and useful for decision support.
Methods
The core contribution is the introduction of a "bias towards relevance" for plan recognition. This bias is operationalized by leveraging the planning knowledge within the TraumAID 2.0 system. TraumAID generates plans based on a goal-directed approach, considering diagnostic and therapeutic needs.
The plan recognition algorithm, TraumaTIQ, uses this planning knowledge in two phases:
- Goal Relevance Ranking: Potential explanatory goals are ranked on a four-level scale:
- Relevant: Goals currently being pursued by TraumAID.
- Potentially Relevant: Goals that are part of an active diagnostic strategy, even if not fully supported by evidence yet.
- Previously Relevant: Goals that were once relevant but are no longer.
-
Irrelevant: All other goals. The algorithm prioritizes explanations involving higher-ranked goals.
-
Procedure Evidence Ranking: For a given action, potential procedures that could explain it are ranked based on how well they are supported by other observed actions. This ranking also uses a four-level scale:
- Completed Procedures: All actions within the procedure have been ordered.
- Partially Completed Procedures: Some actions have been ordered.
- Relevant Procedures: Procedures currently in TraumAID's plan.
- All Other Procedures. The algorithm greedily selects explanations from the highest-ranking relevant goals and the most evident procedures. This approach avoids an exponential search by grouping explanations and prioritizing relevance.
Datasets
The paper describes an evaluation performed on "management plans from the 97 actual trauma cases from the Medical College of Pennsylvania used in the retrospective validation of TraumAID." This appears to be a real-world dataset of clinical cases, not a newly introduced benchmark.
Key Insights and Ideas
- Goal Minimization is Insufficient: The paper convincingly argues that minimizing the number of goals is a flawed bias in complex, multi-goal domains like trauma management.
- Relevance as a Guiding Principle: Contextual relevance, derived from a sophisticated planning system, is a more appropriate and powerful bias for plan recognition in these domains.
- Four-Level Relevance Scale: The proposed four-level scale for both goal and procedure relevance provides a structured way to quantify and prioritize explanations.
- "Benefit of the Doubt" Policy: TraumaTIQ gives physicians the benefit of the doubt, assuming their actions are explainable within the system's understanding of relevant goals, which is crucial for practical clinical decision support.
- Handling Sub-optimal Plans: The bias allows for the interpretation of actions that might be part of a less preferred but still valid procedure, or even actions that are slightly ahead of the system's current evidence threshold.
Key Results
- The TraumaTIQ plan recognizer, implemented in Lucid Common Lisp, processed 584 actions in an average of 0.023 CPU seconds per action on a Sun 4. This demonstrates computational efficiency.
- In an evaluation on 97 trauma cases, 234 out of 584 actions were not part of TraumAID's plan at the time they were ordered.
- 15 of these were explained by goals currently in TraumAID's relevant goal set.
- 69 were explained by potentially relevant goals.
- The remaining 148 actions could not be explained by relevant or potentially relevant goals.
- The paper suggests that the 148 unexplained actions might be due to the need for a more complete goal abstraction hierarchy or the inclusion of goals not currently in TraumAID's knowledge base.
Limitations and Future Work
- Incomplete Goal Hierarchy: The authors acknowledge that the current knowledge base, optimized for plan generation, only includes goals that can be directly operationalized. A more complete plan recognition system would require a richer goal hierarchy, including goals that are not immediately actionable.
- Goals Outside System Scope: Some goals physicians pursue might be outside the current scope of TraumAID's knowledge base, necessitating their inclusion for comprehensive recognition.
- Broad Diagnostic Tests: The paper notes that many broad diagnostic tests are difficult to explain because physicians may order them without a specific, immediate goal in mind. This points to a need for more sophisticated reasoning about uncertainty and probabilistic goal activation.
- Future Work: The authors suggest that a more complete plan recognition system would require addressing these limitations by expanding the knowledge base and developing more advanced methods for handling uncertainty and broad diagnostic actions.
Summary updated