SOHO 2025 | Opportunities and challenges for real-time AI decision support in transplant care

In my view, there are many opportunities for the patients undergoing allogeneic stem cell transplantation, in particular in the terms of monitoring and risk prediction. We know already from previous studies that there’s a strong center effect and that larger centers provide substantially better care to patients and that’s why we try to treat as many patients as possible in centers. And I think this is very well demonstrated and this is something that we appreciate. However, there’s still a certain proportion of patients for several reasons that are treated also in smaller institutions and smaller centers. And this can be geographically motivated, such as patients who have difficulty also in accessing other centers but this also has other reasons such as funding and this is not only a US-based problem but also a problem seen globally because we have centers of different sizes and my view is that decision support systems but also patient empowerment techniques can really be helpful in that part. We have experimental devices ongoing for the monitoring of patients. There have been previous publications that have shown, for instance, that patients who have been continuously monitored can be detected for infectious complications upfront before they develop symptoms or before they develop abnormalities also in the laboratory range that would drive physicians to take action. So this is something where we can have warning and support systems. There are other possibilities also to monitor the patients in order to keep them in a steady health status within this challenging situation after treatment. And this is just in order to summarize what we have on the terms of monitoring and surveillance. For classification of complications, we have been working on a graft versus host disease classifier. There are other groups who are working on the classification of chronic graft versus host disease even utilizing images of the skin so what I just want to emphasize is that there’s a multitude of works going on also in this domain. For the challenges now speaking we have a challenge still with data integration because AI and I mean well-running AI is building on a good structure a very solid hospital data structure and also on principles of data integration and although we have the possibilities we have the common language also to speak between different institutions and we need to have the power of the servers and those institutions ready in order to run the devices. And we need also to have the devices connected with an intimate data structure that is enabling to contact and to derive this information within real time. Because if we have long delays between the accessibility of the data and then also the execution of the models that are predictive, we may have the answers not within the desired timeframe that we have. Another issue that we are still experiencing, and that’s also why the models often stay at the experimental level, is that they become developed and tested within individual institutions but not beyond. So we have the risk that those models are overfitted to a specific use case within those institutions and not so well deployable to other centers. Federated learning can be one answer to address this challenge specifically but the implementation of those models and also the willingness to advance further is still within certain limits. Funding is certainly also one issue because ultimately those devices when they want to become implemented in routine care everywhere need to go through quite rigorous validation process and also accommodation process by the federal agencies such as FDA or EMA. And one solution out of this dilemma in order to accelerate this process is that some institutions and this being said also the caveat of having the institutional limitation when you say, well, you just have an institution running its own device, but it has the possibility to use a so-called hospital exemption in order to employ devices that are not publicly available within the research purpose within the own institution.

This transcript is AI-generated. While we strive for accuracy, please verify this copy with the video.