Multi-signal updating of process models:
the case of Living Guidelines in medical care

Hajo A. Reijers, Annette ten Teije, Rinke Hoekstra

What is the problem & why is it important?

Medical guidelines are deployed to ensure that clinical practice is maximally effective while risks are being minimized. Such guidelines sanction a preferential process or organizational routine to accomplish a medical task in accordance with a fixed set of assumptions, in particular the clinical evidence at the time of their creation.

Ideally, medical guidelines stay in sync with the latest insights from the medical literature. However, new research is published at an astonishing rate - PubMed grows with 2000 papers per day - while it typically takes at least  three years before a single guideline is updated [1]. Common practice relies heavily on Randomized Controlled Trials and committee-based consensus-building.

The question that interests us is: how can we enable a faster cycle of guideline updating and arrive at “living guidelines”, i.e. evidence is automatically collected and interpreted to modify the preferred way of carrying out diagnosis and treatment activities. Guidelines that stay in sync with the medical state-of-the-art can contribute to the well-being of numerous patients and contribute to the economic viability of healthcare systems.

How will we solve it?

The innovative idea behind this proposal is that the increasing availability of various types of data provides opportunities for data-centric approaches to the guideline updating problem, as compared to the slow, labor-intensive process that exists today. In particular, it seems promising to exploit event data on the actual execution of medical processes and open data from a variety of sources.

Initially, we will focus on the following signals to initiate an update of a medical guideline:

1. The identification of a conflict between the prescribed process model (the guideline) and the observed process model (event data);
2. The observation of unacceptably poor results of the prescribed process/guideline (quality indicators);
3. The occurrence of new knowledge in the scientific literature (open data).

To establish the detection and processing of each of these signals, we aim to develop:

1. Analysis techniques to compare the observed behavior of a medical process with respect to the expected/preferred execution as specified in the guideline. The area of process mining provides a viable basis for such techniques;
2. Dependency analysis techniques to identify which parts of the guideline and/or mined process influence the score of a particular quality indicator;
3. Techniques to identify relevant literature to be used in updating  the prescribed model (guideline) based on: (i) new available literature, (ii) discrepancy among guideline and observed behaviour (results from 1), (iii) quality indicator and the related part of the guideline/observed behaviour (results from 2).

Our goals are very challenging. First of all, each of the foreseen steps are knowledge-intensive. Additionally, we know from earlier work that there is a partial disconnect between the terminology and level of description used in guidelines versus clinical data, indicators, and the medical literature [2].  A further challenge is the comparison between normative, abstract, vague and declarative process models expressed in guidelines on the one hand and the concrete instantiation of processes in clinical practice on the other.

The large body of knowledge available in the Linked Open Data cloud will be used as background knowledge to support these steps. Recent work by Zamborlini et al. has shown that linked data helps to detect interactions between guidelines [3]. We will apply this insight to build a bridge from guidelines to mined models, indicators and data, but also to bootstrap process mining (by improving the comparability of activities). We will investigate different methods for mapping mined process models onto guidelines (KR-based, rule-based or based on graph structure).

What do we need to make this succeed?

Much of the data that is required for this project is already available to us at this point. We have ensured access to data from a specific medical domain through our existing contacts with practicing dermatologists. Specifically, this pertains the following data:

• Medical guidelines for dermatology (old and new versions), publically available at http://www.huidziekten.nl/richtlijnen/richtlijnen.htm;
• Quality indicators, both definitions and historically observed values. We will use the indicators defined by the Dutch Health-care Inspectorate (Indicatorenset Melanoom van de huid), as well as a larger set of dermatology indicators from the US national quality measures clearinghouse;
• Hospital data for process mining, through our contacts with medical practitioners with whom we have worked before, see e.g. [4,5];
• Medical Linked Open Data sets and vocabularies such as DrugBank, Sider, AERS, LinkedCT, DIKB, DermBase, UMLS, Snomed, MESH. All of these are publically available, and many of these were interlinked by ourselves in earlier projects such as LarKC, OpenPHACTS, and COMMIT.

Who will we be looking for?

To fill the PhD position for this project, we will be looking for a candidate with either a background in medical informatics or a computer scientist with an outspoken interest in the medical domain. Since the medical guidelines in question are available in Dutch, knowledge of this language is a very strong benefit for a successful execution of the project. Experience with data analysis techniques, in particular process mining, is another advantage.  

References

[1] Shekell, P., Eccles, M. P., Grimshaw, J. M., & Woolf, S. H. (2001). When should clinical guidelines be updated? British Medical Journal, 323(7305), 155.

[2] Marcos, M., Berger, G., Van Harmelen, F., ten Teije, A., Roomans, H., & Miksch, S. (2001). Using critiquing for improving medical protocols: harder than it seems. In Artificial Intelligence in Medicine (pp. 431-442). Springer Berlin Heidelberg.

[3] Zamborlini, V., da Silveira, M., Pruski, C., ten Teije, A., & van Harmelen, F. (2015). Analyzing recommendations interactions in clinical guidelines. In Artificial Intelligence in Medicine (pp. 317-326). Springer Berlin Heidelberg.

[4] Van der Geer, S., Reijers, H. A., van Tuijl, H. F., de Vries, H., & Krekels, G. A. (2010). Need for a new skin cancer management strategy. Archives of Dermatology, 146(3), 332-336.

[5] Reijers, H. A., Russell, N., Van der Geer, S., & Krekels, G. A. (2009). Workflow for healthcare: A methodology for realizing flexible medical treatment processes. In Business Process Management Workshops (pp. 593-604). Springer Berlin Heidelberg.