Grit NickelTomorrow’s healthcare is increasingly data-driven. Collecting robust and reliable medical data is essential for making informed decisions at every step of the medical care process. Artificial intelligence (AI) plays a vital role, enabling the analysis of large datasets to identify patterns and correlations. These patterns and correlations serve as the foundation for developing predictive models that associate specific symptoms or diseases – which assists medical professionals in making well-informed choices.
We talked to Dr. Nadine Lang-Richter, Group Leader of Medical Data Analysis at Fraunhofer IIS. She emphasizes that her work encompasses more than just the pursuit of high-quality data and sophisticated algorithms; it also involves understanding how to effectively apply this knowledge to enhance real-world healthcare outcomes.
Nadine, your expertise includes the analysis of multimodal medical data and the development of AI-based longitudinal analyses and predictive models. That’s a mouthful … Can you explain what it is that you’re doing?
Nadine Lang-Richter: Any patient with a health issue has a multitude of parameters that can be analyzed. For instance, we can use ECG to monitor their heart activity, we can measure their breathing rates, track skin conductance, and gather several other physiological signals. We look at all this data combined, which is what we mean by “multimodal.” We fuse the data with metadata, such as gender, age, height, weight, and with medical history data to get information as accurate as possible about the patient’s condition across all modalities.
Gathering all those parameters and connecting them is very time consuming for medical professionals, when it has to be done during routine clinical practice. It’s also difficult to summarize everything and keep track of it all. And this is where we come in: We build the technology that performs multimodal assessments of the patient’s condition. The “longitudinal” part means we analyze this data over longer periods for each patient. By gathering a lot of information from one patient and by comparing it to a peer-group – a group that has a similar set of symptoms and a similar age and gender structure – we can track how the patient’s condition changes over time. This helps us predict the most likely course of the disease for this patient in the coming months, which in turn allows doctors to provide more personalized treatments.
We are working with medical data, so we can only work with black boxes to a very limited extent.
Could you provide more details on how you gather longitudinal medical data? Or are you using pre-existing data?
Ideally, we start with existing data from previous doctor visits or hospital stays. However, if that information is limited or unavailable, we collect data ourselves. This involves providing patients with wearables or trackers, or bringing them into one of our laboratories for measurements. For example, we have patients who follow a gait analysis protocol while simultaneously recording an ECG. In a best case scenario, we combine the data already collected with the new information we gather, to give us a more comprehensive view of the patient’s condition.
During gait analysis, passive reflective markers are attached to the patient’s body, allowing cameras to capture their movements in real-time and reconstruct a three-dimensional model (© Fraunhofer IIS).
Please walk me through the process: You obtain the data, analyze it, and train an AI – what happens next with the information you extract from the data and the AIs you developed?
Well first, once we have the data, there is a lot of data management that needs to be done. We need to check if the data is synchronized, and usually it’s not, so we have to fix that. Then, the data might come in different formats. We need to optimize it and bring the data into formats the AI can later work with. We also need high-quality data storage to allow for organized access to this data.
AI is the central technology in data analysis, data fusion, and prediction, which can be broken down into three steps. First, in data analysis, AI models are applied to both medical history data and newly collected data, depending on the specific signals involved. Next, during signal fusion, AI fusion models examine the co-dependence and connections among different data sets. Finally, for prediction, meaning the forecasting of the data, AI is used to analyze the past, and develop it into a possible scenario for the proceeding of the health status of the patient future.
To ensure our networks and models are ready for immediate use in medical settings, we prioritize interpretability and explainability, while being conform to GDPR and the EU-AI Act. In the end, we are working with medical data, so we can only work with black boxes to a very limited extent. We make sure that the output of our models reflects medical parameters and help with the interpretation. As a team, we review the output, focusing on its meaning, the data the network relies on, and the input data used. We have a large network of doctors and clinics that are always consulted and involved in the process. That way, we ensure that the information is credible and valuable for the doctor, their patient, and people close to the patient.
Multimodal data acquisition in the driving simulator at Fraunhofer IIS: Investigating cognitive load is a key aspect of automotive health (© Fraunhofer IIS).
Can you share an example of a project and how it benefited patients or medical professionals?
We currently have an R&D project with Sonovum, a company specialized in medical devices for non-invasive brain monitoring. They have developed a device, essentially a headband, that can estimate intracranial pressure using ultrasound. By interpreting these signals, we can determine if the pressure is normal or elevated, which is crucial for stroke and trauma patients.
When people arrive at the clinic with a brain injury, scans might show trauma, but it’s often unclear whether the intracranial pressure is elevated. Until now, the only way to measure this is by inserting a probe into the patient’s skull. When Sonovum’s device detects elevated intracranial pressure, a probe will still be inserted for confirmation, and countermeasures will be taken immediately. However, this development also benefits patients who do not need the probe because their pressure is normal, avoiding an unnecessary invasive procedure that carries risks.
We support the development and training of the AI models to differentiate between elevated and normal intracranial pressure data. The key challenge is understanding the physiological significance of these signals. We’re working with proprietary signals rather than standard ones where the significance is already established. Instead, a lot of work has to be put in to understand what these factors indicate and how they relate to elevated or normal intracranial pressure.
Our vision is to develop solutions that help doctors and medical professionals to better and faster assess a patient’s condition.
That’s a great example of how your developments can affect medical care and patient outcomes. In a nutshell: How does your work contribute to the future of healthcare in general?
Our goal, our vision, is to develop solutions that help doctors and medical professionals to better and faster assess a patient’s condition. This allows them to provide quick, individualized help to their patients. We aim to provide the technology, that allows to combine all the important information from a patient, to improve the accuracy of the assessment of the condition and provide optimal, individualized treatments This reduces pain and suffering for patients, and it eases the workload of doctors, who have many patients to care for at once. Ultimately, it is also beneficial for the healthcare system as a whole if we have less trial and error because that saves costs as well.
Image copyright cover image: Adobe Stock — Ipopba
Tobias Sebastian Zech
Julia Fröhlich
Anna Chiwona
Add comment