gms | German Medical Science

Introduction: Personalized Medicine requires clinical trials to generate evidence [1]. In order to research highly targeted oncologic therapies multi-arm trials become increasingly prevalent [2]. In- and exclusion criteria also see a subsequent rise in complexity but often lack semantic annotation in trial management systems [3]. Manual patient screening for recruitment can therefore be a difficult and time consuming process [4], [5]. One possibility to overcome this challenge are clinical decision support (CDS) systems [6], [7]. Software based on secondary use for data which could discard unsuitable or identify suitable trials, and in a second step even suggest a trial arm for a patient. This could reduce the workload on clinicians by reducing the number of trials to be inspected.

Methods: Our objective was to develop a prototype for a CDS system to pre-screen potential trial participants using machine processable decision models. Decision modelling provides a way of explicitly defining decision logic while potentially aiding implementation and maintenance over time through visual tools. To explore this concept, the prototype was limited to support the decision logic of only one predetermined clinical trial. Two decision stages were proposed: I) determining if a patient meets the general trial criteria and subsequently and II) suggesting a trial arm. As a proof-of-concept, a subset of arms for one trial was supported. After benchmarking different decision modelling approaches like Analytic Hierarchy Process [8] and the Object Management Group Standards for versatility for machine processing capabilities and tooling, a model was created using the Decision Model and Notation (DMN) specification [9], [10]. The prototype was implemented using JAVA Spring Boot [11] with an embedded Camunda DMN engine [12] allowing processing of DMN models. During and after development, the implementation was unit tested using a data set of 415 manually generated test patients.

Results: The model containes 51 decision elements, whose logic was implemented using hierarchical DMN tables, as well as 144 input variables. During testing, the prototype made the expected suggestions for each test case. Leveraging the structure of the DMN model, it was also possible to implement a feedback loop within each step of the decision process.

Discussion: For practical reasons, the amount of user input should be kept to a minimum. Given the cascading nature of the created decision model, only mandatory information which cannot be reutilized from the clinical context, is asked for during each step in the decision process. Information about each decision is provided for transparency and helps the user understand the causes of the result to facilitate a decision-making process. Use of manually generated test cases was a major limiting factor during evaluation impacting test results. Prior to any real-world application, the system must be validated with real world test data to ensure correct and reliable functioning.

Conclusion: This DMN based approaches seems to be a promising method for initial trial screening. The created models can be interpreted by software at runtime. It seems to be sufficient for the use case at hand, where a human is presented with the evaluation result, potentially reducing screening time for trials.

The authors declare that they have no competing interests.

The authors declare that an ethics committee vote is not required.