Die von unseren Teammitgliedern vorgestellten Konferenzbeiträge, Kurzvorträge, Workshops und Poster deckten ein breites Spektrum an Themen ab:
Evidence-Grounded LLM Validation of MIMI-IV ICD Labels
How can we ensure that the evaluation of clinical NLP models truly reflects the patient facts documented in medical narratives rather than administrative artifacts?
One major challenge in automated ICD coding remains the presence of substantial label noise in widely used benchmarks like MIMIC, where many charted diagnosis codes are not text-grounded in the discharge summaries, leading to confounded model training and biased performance estimates.
Our colleague Ahmad Abu Dayeh presented an innovative approach to address this challenge by leveraging Large Language Models (LLMs) to systematically verify and refine dataset labels.
For further details you can read the corresponding conference paper: https://ebooks.iospress.nl/doi/10.3233/SHTI260328
Distribution Shift Analysis in Generalizable Modelling: Intensive Care Time-Series Data
Is there a way to contextualize the performance of medical AI on private internal and external validation sets?
One major challenge in the deployment of medical AI remains performance degradation in clinic-to-clinic scenarios, due to, e.g., differences in patient demographics, measurement devices, and clinical practices, which manifest as a distribution shift.
Our colleague Mayra Elwes presented first explorations towards potential strategies for benchmarking generalizability by incorporating distribution shift measures in her presentation on "Distribution Shift Analysis in Generalizable Modelling: Intensive Care Time-Series Data".
For further details you can read the corresponding conference paper: https://ebooks.iospress.nl/doi/10.3233/SHTI260163
Showcasing PADME as a Distributed Analytics Platform in a Real-world Healthcare Setting
How can we seamlessly scale advanced health data analytics across independent clinical institutions while keeping patient data entirely secure behind local firewalls?
One major challenge in the deployment of distributed medical AI remains the difficulty of demonstrating that decentralized architectures can move past isolated, theoretical proof-of-concepts and smoothly handle the practical, administrative, and multi-modal data complexities of real-world clinical environments.
Our colleague Mehrshad Jaberansary showcased our PADME platform in the live demonstration to prove exactly how secure analytical pipelines can visit decentralized hospitals and compute insights on native data without it ever leaving its point of origin.
Implementing a Governance Framework for Federated Learning
How can we seamlessly bridge the gap between collaborative multi-institutional AI research and the highly rigid, secure compliance barriers of individual hospital networks?
One major challenge in the deployment of federated learning in healthcare remains its limited real-world application within the highly secure hospital domain, primarily due to the complex regulatory and administrative hurdles of managing secondary data use across distinct clinical institutions without violating structural security protocols.
Our colleague Ana Grönke presented the operational governance framework to discover how it can be adapted to smoothly facilitate the AI-driven secondary use of clinical routine data in a federated setting.
For further details you can read the corresponding conference paper: https://doi.org/10.3233/SHTI260418
Traceability in Federated Learning: A Scoping Review
How can we achieve transparency collaborative healthcare AI when the underlying data must remain private and distributed?
One major challenge in federated learning (FL) within healthcare remains the difficulty of ensuring traceability as one aspect of transparency, as the decentralized nature of federated learning makes it complex to monitor, verify, and trace processes, models, users or data.
Our colleague Kim Tang presented first findings of the scoping review to map the current state on challenges and solutions regarding traceability in federated learning settings.
For further details you can read the corresponding conference contribution: https://ebooks.iospress.nl/doi/10.3233/SHTI260281
