The penicillin of our day

^{Web special Fraunhofer magazine 1.2023}

Every minute counts when caring for severely injured persons. They are treated in the trauma room, which is located right at the heart of the emergency department and equipped with cutting-edge technology, respirators and powerful full-body CT scanners. The trauma room team need to determine any possible fractures or injuries to the brain or spinal cord as quickly as possible. Did the patient fall from a height of more than 3 meters? Were they in a bicycle crash? Are they diabetic? Every single detail affects crucial decisions, from the specialists should be brought in to help to the treatment steps taken – the also determine the order in which these actions are taken. This information sometimes makes the difference between life and death, but as it stands, it is only passed on orally by emergency service personnel. It is almost never recorded in any structured manner, creating a huge risk of information going missing. 

Smart assistance for decision-making

Researchers at the Fraunhofer Institute for Intelligent Analysis and Information Systems IAIS are working to change this alongside their partners in the TraumAInterfaces project. They want to build an AI-driven speech recognition system that will automatically record, analyze and structure the information exchanged during the handover of severely injured patients. “Our AI does not make any decisions,” emphasizes Dario Antweiler, head of the Healthcare Analytics business unit at Fraunhofer IAIS. Its sole purpose is to support the trauma room team’s decision-making processes as effectively as possible. A structured record of all the necessary information and the course of treatment is also valuable for documenting the case at later stages. “Producing this documentation wastes a lot of valuable time that doctors could be spending on patients instead,” Mr. Antweiler points out.

He and his team spent days and days visiting various different hospital departments, observing work and administrative procedures, interviewing doctors and talking to hospital experts. “What we focused on above all was, what IT systems are they using? How do they communicate? Where is data generated and how is it used?” Next, the data scientists developed more than 40 use cases for deploying AI in such as way as to reduce the workload on medical staff at hospitals and improve the quality of treatment. “Patients are getting older and older, and more and more numerous, and the personnel shortage is becoming more serious. In the future, hospitals simply won’t be able to get by without AI,” Mr. Antweiler asserts.

Saving precious time with AI

Artificial intelligence holds enormous potential for hospitals. AI-driven systems could establish statistical correlations based on hundreds of thousands of previous cases, allowing them to predict possible complications and risks at an early stage, for example. When combined with doctors’ experiences, this information could be used to adjust treatment strategies accordingly, and improve the patients’ chances of survival. AI can also help estimate the risks of a surgery in a more accurate way. Doctors could use “literature mining” to quickly find up-to-date information about the best treatment options – which would be a particular advantage in the complex, interdisciplinary, time-sensitive process of treating the severely injured. In literature mining, the AI analyzes large quantities of specialist medical literature, which is automatically updated on a regular basis. Not only is this process far quicker than searching online databases in the conventional way, but it’s also much more accurate. AI can run semantic searches, meaning that rather than just searching for specific terms, it also searches for related content – and it can filter and classify the results, ensuring that a critical care unit is shown different articles from a regular unit, for example. “These days, AI is very good at extracting information from texts, which is something the healthcare sector desperately needs. Almost all information is in text form, from diagnostic findings and doctors’ letters to documentation. Producing, reading and analyzing these texts takes up incredible amounts of time,” explains Mr. Antweiler. However, clinicians simply do not have the time to study documents like a thick patient file containing 20 years of medical history, so any important information it may contain can easily go unnoticed. 

Producing doctors’ letters quickly – at last!

Now, Mr. Antweiler and his team are working in the SmartHospital.NRW project to automatically extract important details on diagnoses, prior treatments and allergies via AI and present it in a clear, structured form for medical personnel to use. Starting in 2024, AI-driven text mining well be deployed at University Hospital Essen, one of the project partners and a trailblazer in the field of digital transformation in Germany. The hospital is also set to put the AI-based speech recognition system to the test in the near future, along with another Fraunhofer IAIS innovation, the doctors’ letter generator. These documents, which are issued to every patient upon leaving the hospital, serve as the primary means of communication between hospitals and doctors’ offices. However, they currently take doctors an average of three hours to produce. The letters contain details such as the patient’s case history and suspected diagnosis assigned upon their admission to inpatient care, as well as any drugs that have been administered and any treatment steps that were taken in the hospital. Until now, medical personnel have had to go through the laborious process of collecting all this information from different IT systems. In the future, however, an AI program will automatically extract this information and insert it into the doctor’s letter. Only the epicrisis, that is, the summary of the overall hospital stay, the conclusions to be drawn and the recommended treatment will still have to be drafted by the doctor for the final full text.

Prof. Jochen Werner, medical director and head of the executive board at University Hospital Essen, wants to make his clinic the first “smart hospital” in Germany – a mission that is already well underway, with the support of Fraunhofer IAIS. “Data is the penicillin of our day,” Prof. Werner announces. He uses his YouTube channel to promote digitalization in medicine, with a view to allaying the fears of medical personnel and patients alike.

University Hospital Essen started the process of converting its hospital information system – which had previously been used primarily as a billing system for medical services – into a data management system more than ten years ago. Now, all the available details on a patient and all the data that is generated during their hospital stay, from blood values and body temperatures to medications, is stored in the Smart Hospital Information System (SHIP). All hospital equipment is connected to this platform, so if a patient’s blood pressure is measured, for example, the device will automatically record this value in the system. An open standard is a prerequisite for this setup, i.e. the software must allow interfacing with SHIP. “An open standard called FHIR is currently being developed in the international healthcare scene. Instead of text fields, this standard works with set codes that can be read anywhere in the world, with meanings like ‘blood pressure decreased’,” explains Mr. Antweiler. 

Optimizing the search for ideal test persons for clinical trials

The problem is that so far, most patient data in Germany has not been recorded in a structured form nor stored in interoperable IT systems that can be connected to each other – this leads to issues in, for example, clinical trials that concern a number of different diseases. Even the process of searching for suitable trial participants is immensely time-consuming, as hospitals and doctors’ offices must review their patient bases manually and compare them against the trial’s catalog of requirements. “At the start of the trial, researchers simply estimate how many patients should be involved. It’s not uncommon for these estimates to miss the mark. If the trial fails to reach the necessary size, then in the worst case scenario, it could fail, which can cost millions of euros,” explains Sabine Kugler, a data scientist at Fraunhofer IAIS.

She is currently working on project PARIS, where she extracts the information needed for clinical trials from running texts and compares it with the requirement profiles. In an initial use case, she took the files of patients with psoriatic arthritis as a data basis. Approximately one third of the people that suffer from psoriasis go on to develop this particular form of the condition, which is associated with inflammation of musculoskeletal structures. Researchers at the Fraunhofer Institute for Translational Medicine and Pharmacology ITMP worked with rheumatologists from Frankfurt university hospital to tag items of information that count as important selection criteria for clinical trials, such as the date of the patient’s diagnosis, the intensity and duration of their pain, and the stiffness they experience in the mornings. “We use this as a basis for training our AI model. It uses the sample data we feed it to learn and even formulate rules independently. The more data we give it, the better the results,” explains Ms. Kugler. The information it extracts will automatically be given a structure later on and saved in an easily searchable database. Patients can be filtered based on certain criteria via a search query. “Our AI model can easily be adjusted to other rheumatic conditions. We need less data for that, because the model has already learned many important characteristics of rheumatism. However, it would have to be completely retrained from the ground up for other diseases, such as cancer or diabetes,” Ms. Kugler relates. 

Unravelling the complex interaction of possible disease factors

But the Fraunhofer researchers are going even further. They also want to use large quantities of data to improve the processes of diagnosing and treating psoriatic arthritis (PsA). Since July 2021, they and 26 partners from Europe, the USA and Canada have been working on project HIPPOCRATES, an initiative aimed at using data analysis to gain a deeper understanding of the complex interplay of a range of different factors involved in the emergence of PsA. After all, the inconsistent symptoms and the highly varied progression associated with PsA are precisely what make it so difficult to recognize and treat at an early stage. The condition often involves inflammation of the synovial membrane on the patient’s hands and feet, but their tendons, eyes or intestines may also be affected. “This huge variation in the clinical symptoms means that the illness is diagnosed too late in most cases, which then significantly reduces the chance of the patient receiving effective treatment. This is why we need diagnostic tools that will allow us to make reliable diagnoses in practice,” explains Dr. Michaela Köhm, a rheumatologist and leader of a research group at the Frankfurt-based Fraunhofer ITMP’s Clinical Research division that collaborates closely with the university hospital.

Envying the Scandinavians

HIPPOCRATES is combining the datasets and cohorts from the largest European PsA studies, so that an AI program can search for patterns that consistently occur across patient characteristics recorded in clinical settings, imaging procedures and molecular testing. Through this, the researchers hope to develop an accurate criteria catalog for more effectively predicting the disease and making prognoses. “In the clinical field, you would count yourself lucky if you had a dataset of just 1,000 patients,” relates Sina Mackay, a data scientist at Fraunhofer IAIS. “But in HIPPOCRATES, we have many times that. Converting all this data into a compatible format is quite a feat, but it will soon pay off.” That is not to say that the only issue with the data sets will be their uniformity: they will also likely leave much to be desired in terms of completeness. “As a result, we are looking for a machine learning method that deals well with missing values,” explains Ms. Mackay.

In fact, Ms. Mackay, Dr. Köhm, Ms. Kugler and Mr. Antweiler feel some envy toward Scandinavia, where researchers can access much larger volumes of patient data for projects like HIPPOCRATES. For example, Finland has digitized all of the health data that it has collected since 1960, making the anonymized medical information available to researchers. Estonia is another target for the green-eyed monster: it also uses a central repository to collect and store its health data in a central repository in anonymized form, providing this to scientists upon request. “Here in Germany, we urgently need to find ways of facilitating medical research based on large volumes of health data while complying with legal requirements,” insists Dr. Köhm. “Otherwise, we are missing opportunities to spot patterns that could help us draw conclusions about individual risks for illnesses, comorbidity or medication side effects; not to mention, we are losing the chance to keep pace with international research.” We can only understand individual cases by studying the multitudes.