1 Applying intention-based guidelines for critiquing Robert-Jan Sips, Loes Braun, and Nico Roos Department of Computer Science, Maastricht University, P.O.Box 616, 6200 MD Maastricht.
2 Contents Introduction. Intention-based matching. Experiments. Results. Conclusions. Further Research.
3 Introduction Intention-based matching Experiments Results Conclusions Further Research -Standardisation of care. -Current development: Evidence-based. -Proven improvement of care. but -Physicians tend to reject „cookbook-medicine“. -Not flexible concerning deviations. Medical Guidelines
4 Introduction Intention-based matching Experiments Results Conclusions Further Research -(Expert) system providing feedback on performed actions. -Guiding the physician in a subtle manner. but -Difficult to adapt to new developments. -Current systems rely on user interaction. Expert Critiquing Systems
5 Introduction Intention-based matching Experiments Results Conclusions Further Research -Proposal: Combine expert critiquing and medical guidelines. + = -Prerequisite: Matching a physician‘s actions (reported in an EPR) and those prescribed in a medical guideline. (No user interaction). The best of two worlds
6 Introduction Intention-based matching Experiments Results Conclusions Further Research Matching Previous research learns: 1.Physicians do not follow a guidelines exact actions. (Van der Lei (1991)). 2.Solution: Match intentions (Advani et al. (1998)). 3.Differences in intentions reported by a physician and modeled in a guideline (Marcos et al.(2001)).
7 Introduction Intention-based matching Experiments Results Conclusions Further Research Observation: -2 types of intentions: 1.High-level intentions. Diagnosis and treatment goals. E.g. 2.Low-level intentions. Application independent intentions of clinical interventions. E.g. Low-level intentions. Described in standard literature (e.g. Merck Manual, pharmacotherapeutical compass). Intentions
8 Introduction Intention-based matching Experiments Results Conclusions Further Research 1.Each high-level intention can be described by a set of high- level and low-level intentions. 2.Each high-level intention can be described by a set of low- level intentions. Therefore The most general high-level intention in a guideline can be replaced by one or more sequences of low-level intentions: the guideline execution. Medical Guidelines
9 Introduction Intention-based matching Experiments Results Conclusions Further Research -Measure of similarity between the low-level intentions performed by a physician and a guideline execution. -Informal: |physician actions in the execution| - |actions in the execution not performed by the physician| Distance to guideline
10 Introduction Intention-based matching Experiments Results Conclusions Further Research -Asbru modeled guideline (obtained from Marcos et all). -Case interpretations from two pediatricians. -Case interpretations entered in EPR in 3 ways. -Normal (as reported by the pediatrician). -Consultation basis (3 actions in arbitrary order per consult). -Reversed order (worst-case scenario). Experiments
11 Introduction Intention-based matching Experiments Results Conclusions Further Research -Normal sequence: -Average of 69,4% correct over all actions. -Without actions outside the guidelines scope: 80,1% correct. -Without actions outside the guidelines scope and correction for error in the guideline 95,8% correct. -Better performance on long sequences than on short. -No significant difference in performance between the two physicians after discarding a case outside the guideline‘s scope (Liver infection as cause). -No significant difference in performance on sequences in different orders. Results
12 Introduction Intention-based matching Experiments Results Conclusions Further Research -RBE and backwards -No significiant difference in the performance on sequences in a different order. Results
13 Introduction Intention-based matching Experiments Results Conclusions Further Research 1.Algorithm performs adequately. 2.Our results support the claim that physicians follow a guideline‘s intentions. 3.Our results indicate that there is no difference in performance on different treatment styles. Conclusions
14 Introduction Intention-based matching Experiments Results Conclusions Further Research -Test our algorithm more extensively. Prove performance. -Different measures for matching. E.g. Use Temporal data. -Expand our algorithm to match on multiple guidelines. Changing treatment goals. -Using real-life patient records. Terminology. -Effect on the treatment process. Does this way of critiquing improve care? Further Research