Introduction: Widespread adoption of AI for medical decision-making is still hindered due to ethical and safety-related concerns. For AI-based decision support systems in healthcare settings, it is paramount to be reliable and trustworthy. Common deep learning approaches, however, have the tendency towards overconfidence when faced with unfamiliar or changing conditions. Inappropriate extrapolation beyond well-supported scenarios may have dire consequences highlighting the importance of the reliable estimation of local knowledge uncertainty and its communication to the end user. Materials and Methods: While neural network ensembles (ENNs) have been heralded as a potential solution to these issues for many years, deep learning methods, specifically modeling the amount of knowledge, promise more principled and reliable behavior. This study compares their reliability in clinical applications. We centered our analysis on experiments with low-dimensional toy datasets and the exemplary case study of mortality prediction for intensive care unit hospitalizations using Electronic Health Records (EHRs) from the MIMIC3 study. For predictions on the EHR time series, Encoder-Only Transformer models were employed. Knowledge uncertainty estimation is achieved with both ensemble and Spectral Normalized Neural Gaussian Process (SNGP) variants of the common Transformer model. We designed two datasets to test their reliability in detecting token level and more subtle discrepancies both for toy datasets and an EHR dataset. Results: While both SNGP and ENN model variants achieve similar prediction performance (AUROC: $\approx 0.85$, AUPRC: $\approx 0.52$ for in-hospital mortality prediction from a selected MIMIC3 benchmark), the former demonstrates improved capabilities to quantify knowledge uncertainty for individual samples/patients. Discussion/Conclusions: Methods including a knowledge model, such as SNGP, offer superior uncertainty estimation compared to traditional stochastic deep learning, leading to more trustworthy and safe clinical decision support.

Acute patient treatment can heavily profit from AI-based assistive and decision support systems, in terms of improved patient outcome as well as increased efficiency. Yet, only very few applications have been reported because of the limited accessibility of device data due to the lack of adoption of open standards, and the complexity of regulatory/approval requirements for AI-based systems. The fragmentation of data, still being stored in isolated silos, results in limited accessibility for AI in healthcare and machine learning is complicated by the loss of semantics in data conversions. We outline a reference model that addresses the requirements of innovative AI-based research systems as well as the clinical reality. The integration of networked medical devices and Clinical Repositories based on open standards, such as IEEE 11073 SDC and HL7 FHIR, will foster novel assistance and decision support. The reference model will make point-of-care device data available for AI-based approaches. Semantic interoperability between Clinical and Research Repositories will allow correlating patient data, device data, and the patient outcome. Thus, complete workflows in high acuity environments can be analysed. Open semantic interoperability will enable the improvement of patient outcome and the increase of efficiency on a large scale and across clinical applications.

Background: Obtaining large amounts of real patient data involves great efforts and expenses, and processing this data is fraught with data protection concerns. Consequently, data sharing might not always be possible, particularly when large, open science datasets are needed, as for AI development. For such purposes, the generation of realistic synthetic data may be the solution. Our project aimed to generate realistic cancer data with the use case of laryngeal cancer. Methods: We used the open-source software Synthea and programmed an additional module for development, treatment and follow-up for laryngeal cancer by using external, real-world (RW) evidence from guidelines and cancer registries from Germany. To generate an incidence-based cohort view, we randomly drew laryngeal cancer cases from the simulated population and deceased persons, stratified by the real-world age and sex distributions at diagnosis. Results: A module with age- and stage-specific treatment and prognosis for laryngeal cancer was successfully implemented. The synthesized population reflects RW prevalence well, extracting a cohort of 50,000 laryngeal cancer patients. Descriptive data on stage-specific and 5-year overall survival were in accordance with published data. Conclusions: We developed a large cohort of realistic synthetic laryngeal cancer cases with Synthea. Such data can be shared and published open source without data protection issues.

In surgery, the possibilities of fully automatic assistance are limited. A valuable automated assistance can be implemented by realizing context-aware behaviour of control consoles in integrated operating rooms (ORs). We developed an intraoperative pipeline for surgical context interpretation and an integrated console. Based on that, we propose a method for the context-aware ranking of configuration profiles, which facilitate the handling of devices during endoscopic interventions. The method is evaluated with twenty-four simulated endoscopic ear-nosethroat (ENT) workflows. The results indicate that a continuous automatic ranking based on the surgical situation, can significantly reduce manual user interaction while the surgeon’s control is preserved.