gms | German Medical Science

Introduction: Reusing data generated during clinical care is a crucial factor for clinical use cases where information is fragmented, abundant, hard to manage and growing rapidly in volume. In precision oncology, molecular profiling is established practice and an important input parameter for molecular tumor boards (MTB) where personalized treatment recommendations are discussed [1]. Clinical Decision Support Systems (CDSS) leverage data from the case database in order to reduce the burden on clinicians finding suitable treatment options, which also includes identifying possibilities for enrollment in clinical trials [2].

Methods: We propose Knowledge Artifacts (KAs), an abstract entity holding case information originating in the clinical care context of an MTB. KAs consist of two main segments: i) case characterization and ii) related information. To accommodate the MTB use case, we narrowed down case characterization to diagnosis and relevant mutated genes. For data extraction we considered literature and clinical trials mentioned in free text comments to the MTB case written by the clinical team consisting of biologists, human geneticists and bio informaticians. Data extraction from comments was implemented using a naÏve pattern matching approach. Deduplication of literature was handled using a community based, open-source literature translation service. The possibility to query KAs is provided via a persistent storage and application programming interface (API).

Results: We analyzed 987 user provided comments of 1530 cases of an MTB in southern Germany and created 448 KAs. Pattern based matching returned 897 Document Object Identifiers and 530 clinical trials. PubMed links from comments were translated to DOIs, deduplicated on a per case base and added to the KA.

Discussion: The results show that information extraction from free text comments with a naÏve approach is possible, however it imposes limitations such as edge cases that cannot be matched leading to incomplete information extraction. Undetected contextual nuances within user comments could also lead to misleading KAs, which makes validation a necessity. AI based methods of information extraction could account for contextual information [3] but would still require manual validation. KAs provide a starting point for tracking the evolution of clinical decision making over time due to the persistence of KA repositories. Through the inherent anonymization of KAs, it is possible to generate a common ground for MTBs by structuring user input and using interoperable structures, making this knowledge accessible.

Conclusion: This self-learning prototypical implementation shows feasibility to operationalize knowledge contained in MTB case preparation data. It bears the potential to decrease MTB preparation time while increasing patient outcomes by delivering information within the clinicians’ workflow as CDSS. In the future, enhancing this approach with machine learning algorithms in the data extraction workflow could amplify its potential by generating robust evidence through advanced pattern recognition and analysis [4].

The authors declare that they have no competing interests.

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