Comparing strategies to generate experience-based clinical process recommendations that leverage similarity to historic data
| Autor: | Frederik Gailly, Geert Poels, Diederik Van Sassenbroeck, Steven Mertens |
|---|---|
| Rok vydání: | 2019 |
| Předmět: |
Reflection (computer programming)
Operations research Process (engineering) Computer science business.industry 02 engineering and technology Brier score 020204 information systems Similarity (psychology) Information technology management 0202 electrical engineering electronic engineering information engineering Leverage (statistics) 020201 artificial intelligence & image processing Set (psychology) business Communication channel |
| Zdroj: | ICHI |
| DOI: | 10.1109/ichi.2019.8904693 |
| Popis: | Doctors executing flexible and knowledge-intensive healthcare processes are constantly confronted with decisions on what to do next from a large and diverse set of options. Such decision making can be very demanding and even overwhelming, certainly for less experienced doctors. The current IT infrastructure remains lacking in support for decision making. Adding an additional channel for externalizing the knowledge and experience from historic data as guidance would have clear benefits. In this paper, we take the first step towards offering such guidance in the form of clinical process recommendations for the next activity to be executed. The purpose of these recommendations is not to steer the process according to some optimization criteria, but rather to offer a direct reflection of the experience encapsulated in previous executions of the process. Forty strategies were defined to calculate probability estimates for each possible next activity. The recommendations are subsequently generated as a list of possible next activities, sorted according to the highest calculated probability. The strategies were implemented and applied to three different healthcare data sets and evaluated on average accuracy, multi-class brier score, log loss, how consistent the recommendation rankings are, and the required computation time. The results indicate that responded frequency, variable-position activity similarity and combined strategies perform well for this type of processes. |
| Databáze: | OpenAIRE |
| Externí odkaz: |
|