Search

Federated learning (FL) is a privacy preserving approach to learning that overcome issues related to data access, privacy, and security, which represent key challenges in the healthcare sector. FL enables hospitals to collaboratively learn a shared prediction model without moving the data outside their secure infrastructure. To do so, after having sent model updates to a central server, an update aggregation is performed, and the model is sent back to the sites for further training. Although widely applied on neural networks, the deployment of FL architectures is lacking scalability and support for machine learning techniques such as decision tree-based models. The latter, when embedded in FL, suffer from costly encryption techniques applied for sharing sensitive information such as the splitting decisions within the trees. In this work, we focus on predicting hemodynamic instability on ICU patients by enabling distributed gradient boosting in FL. We employ a clinical dataset from 25 hospitals generated based on the Philips eICU database and we design a FL pipeline that supports neural-based boosting models as well as conventional neural networks. This enhancement enables decision tree models in FL, which represent the state-of-the-art approach for classification tasks involving tabular clinical data. Comparable performances in terms of accuracy, precision, recall and F1 score have been reached when detecting hemodynamic instability in FL, and in a centralized setup. In summary, we demonstrate the feasibility of a scalable FL for detecting hemodynamic instability in ICU data, which preserves privacy and holds the deployment benefits of a neural-based architecture.

Introduction/ Background The availability of digital pathology creates opportunities for the adoption of advanced workflow solutions focused on facilitating and improving the current way of working in pathology labs. Workflow-driven applications can help achieve increased efficiency and quality, support collaboration, and provide detailed insights into the lab processes. The implementation of workflow solutions also creates effective means to monitor and measure activities, and to detect and solve issues. Our solution helps improve processes in the pathology lab (both with respect to efficiency and quality) by modeling and optimizing the existing workflows and by incorporating decision models for automatic execution of relevant tasks and path selection in these workflows. Examples of decision models relate to the implementation and automatic execution of protocols, detection of quality issues, and automatic evaluation of tests with image analysis to evaluate the need for pre-ordering additional tests. Aims This work focuses on the modeling and optimization of relevant pathology workflows to enable clinical users to efficiently and effectively leverage the deployed digital pathology solutions for faster diagnosis and better patient outcomes. Next to identifying and addressing bottlenecks in the workflow, we aim to improve performance by enriching the workflows with clinical decision support. Methods We build a workflow-driven application enabling us to support and propose optimizations for pathology processes, while leveraging the availability of a digital pathology system. We select relevant workflows and identify opportunities for automating tasks and incorporating decision support. The selected pathology processes are represented according to the BPMN standard [BP]. We used the jBPM [jb] workflow suite (compliant with BPMN 2.0) for the modeling and execution of the processes. Programmatic tasks in the workflows are linked to external services executing the logic required by the tasks. Results We proposed a workflow solution enabling the representation of decision models as externalized executable tasks in the process definition. Our approach separates the task implementations from the workflow model, ensuring scalability and allowing for the inclusion of complex decision logic in the workflow execution. In we depict a simplified model of a pathology diagnosis workflow (starting with the digitization of the slides), represented according to the BPMN modeling conventions. The example shows a workflow sequence that automatically orders a HER2 FISH when IHC is borderline according to defined customizable thresholds. The process model integrates an image analysis algorithm that scores images. Based on the score and the thresholds the decision model evaluates the condition and recommends the pre-ordering of an additional test when the score falls between the two thresholds. This leads to faster diagnosis and allows balancing the costs of an additional test versus the overhead of the pathologist by choosing the values of the thresholds., Diagnostic Pathology, Vol 1 No 8 (2016): 13. European Congress on Digital Pathology

To support the efficient execution of post-genomic multi-centric clinical trials in breast cancer we propose a solution that streamlines the assessment of the eligibility of patients for available trials. The assessment of the eligibility of a patient for a trial requires evaluating whether each eligibility criterion is satisfied and is often a time consuming and manual task. The main focus in the literature has been on proposing different methods for modelling and formalizing the eligibility criteria. However the current adoption of these approaches in clinical care is limited. Less effort has been dedicated to the automatic matching of criteria to the patient data managed in clinical care. We address both aspects and propose a scalable, efficient and pragmatic patient screening solution enabling automatic evaluation of eligibility of patients for a relevant set of trials. This covers the flexible formalization of criteria and of other relevant trial metadata and the efficient management of these representations.

Current post-genomic clinical trials in cancer involve the collaboration of several institutions. Multi-centric retrospective analysis requires advanced methods to ensure semantic interoperability. In this scenario, the objective of the EU funded INTEGRATE project, is to provide an infrastructure to share knowledge and data in post-genomic breast cancer clinical trials. This paper presents the process carried out in this project, to bind domain terminologies in the area, such as SNOMED CT, with the HL7 v3 Reference Information Model (RIM). The proposed terminology binding follow the HL7 recommendations, but should also consider important issues such as overlapping concepts and domain terminology coverage. Although there are limitations due to the large heterogeneity of the data in the area, the proposed process has been successfully applied within the context of the INTEGRATE project. An improvement in semantic interoperability of patient data from modern breast cancer clinical trials, aims to enhance the clinical practice in oncology.

Current research in oncology, require the involvement of several institutions participating in clinical trials. Heterogeneities of data formats and models require advanced methods to achieve semantic interoperability and provide sustainable solutions. In this field, the EU funded INTEGRATE project aims to develop the basic knowledge to allow data sharing of data from post-genomic clinical trials on breast cancer. In this paper, we describe the procedure implemented in this project and the required binding between relevant terminologies such as SNOMED CT and an HL7 v3 Reference Information Model (RIM)-based data model. After following theHL7 recommendations, we also describe the main issues of this process and the proposed solution, such as concept overlapping and coverage of the domain terminology. Despite the fact that the data from this domain presents a high level of heterogeneity, the methods and solutions introduced in this paper have been successfully applied within the INTEGRATE project context. Results suggest that the level of semantic interoperability required to manage patient data in modern clinical trials on breast cancer can be achieved with the proposed methodology.

This paper describes the data transformation pipeline defined to support the integration of a new clinical site in a standards-based semantic interoperability environment. The available datasets combined structured and free-text patient data in Dutch, collected in the context of radiation therapy in several cancer types. Our approach aims at both efficiency and data quality. We combine custom-developed scripts, standard tools and manual validation by clinical and knowledge experts. We identified key challenges emerging from the several sources of heterogeneity in our case study (systems, language, data structure, clinical domain) and implemented solutions that we will further generalize for the integration of new sites. We conclude that the required effort for data transformation is manageable which supports the feasibility of our semantic interoperability solution. The achieved semantic interoperability will be leveraged for the deployment and evaluation at the clinical site of applications enabling secondary use of care data for research. This work has been funded by the European Commission through the INTEGRATE (FP7-ICT-2009-6-270253) and EURECA (FP7-ICT-2011-288048) projects.