Dr. Roland Roller is a researcher and project manager in the Speech and Language Technology group at the Deutschen Forschungszentrum für Künstliche Intelligenz (DFKI). At the Berlin location he concentrates on topics related to natural language processing and machine learning, with a great interest in the biomedical and clinical area. #KI-Berlin spoke with Dr. Roller about his research projects MACSS and BigMedilytics, the accessibility of clinical data and texts, and hospital visits of the future.
You are a researcher and project manager in the Speech and Language Technology group at DFKI. Which topics and in which application areas are you researching with machine learning?
In general, we are working in our group on a multitude of questions and application areas; machine learning is always a partial aspect. However, these mainly have a focus on language, especially written language. Our research group is working on topics such as interaction, mobility, news/media and health; my main focus is on health.
We are quite a large group, so the areas of application are very diverse. For example, people are working with chatbots to detect hateful or false messages (keywords: "hate speech" and "fake news"). Others, on the other hand, are working in the area of "simple language", i.e. mapping complex language into a more easily understandable form or summarizing information ("summarization"). I and several of my colleagues are working in the field of information extraction, in other words the recognition and extraction of relevant information from texts as well as their contexts. There are a variety of techniques which are used. In general, these methods make it easier to access information in large amounts of data and/or to enrich ontologies (networks of terms) and knowledge graphs. For example, there are approaches to identify possible side effects from medical publications or social networks.
Since language is very diverse, these fields of application are not necessarily trivial. I am also working on models to identify high-risk patients, for example to detect unwanted hospitalizations at an early stage. Text data also plays a role here (for example clinical notes or doctor's letters), but we also work with a variety of other information which is combined in a model.
Many research groups, start-ups and companies are making great efforts in natural language processing. What about the healthcare sector? Where can the daily work there be made more efficient with new solutions?
If you specifically mention the health sector, I naturally think directly of application in hospitals. In my point of view, a large number of processes can be optimized within a hospital and therefore costs can be reduced. This is presumably due to an improved infrastructure, but also a large number of AI topics. Natural language processing would only be a partial aspect here. Otherwise, digitisation has resulted in the collection of data in hospitals over the last 10-20 years. There is a great deal of potential here in many respects. Searching and finding relevant information occur to me - commercial search engines have no access to this data - for example to find special cases. But also the derivation of knowledge or the identification of special patterns in the data is very interesting. In administration, too, there are approaches to carry out optimised billing for hospitals, for example, with the support of specialist staff, but also with AI. The aim is to invoice everything that was also carried out during treatment. If something is forgotten, the hospital loses money.
Let's talk about your projects: over the past three years, you and your team have been working on the development of a novel, patient-centred eHealth platform. How exactly can one imagine the MACSS (Medical Allround Care Service Solutions) project? What were the results?
The aim of MACSS was to create a patient-centred eHealth platform to connect the physicians at the Charité Transplantation Centre not only with their patients, but also with other attending physicians. One of the problems faced by physicians is that patients only come to the clinic once a quarter. During the rest of the year, they do not always know what is happening to the patients. For example, whether medication changes have been made by other physicians or whether other problems may have occurred. Apart from this, patients often cannot remember exactly what happened maybe two months ago. The Charité physicians have repeatedly emphasized that adherence to treatment is very relevant for the success of the treatment of chronically ill patients. Therefore, the main goal of MACSS was to create a platform via which transplanted kidney patients can better document their daily medication and other vital parameters and share them with their treating physicians via a secure channel. MACSS also takes into account a large number of other aspects, such as the exchange of information between specialists via this platform if the patient so wishes. In this way, the state of health can be better documented with the help of additional information and problems can be dealt with more quickly.
Within the scope of MACSS, the DFKI, like Beuth University, has been working in the field of information extraction from clinical texts. Since these technologies are very domain-dependent and clinical texts are not normally so easily accessible, there are no existing programs and tools for processing these data which can be used - at least not for German. In the English-speaking world there are already various programs that can be used here. In the context of MACSS, DFKI has annotated data and provided basic models for information extraction from German clinical texts. These models can now for example be integrated into more complex applications.
The current difficulty in the German health care system is that data is not brought together centrally. How can resources be pooled - as MACSS is aiming to do - without sensitive patient data being endangered? Does politics need more initiatives in this area?
Pooling resources is a problem, but there are many challenges and difficulties. From my point of view as a scientist, access to data and data quality is a more serious problem. Many scientists, but also companies, work with clinical data and clinical texts. However, each research group often works on its own. There are hardly any common data sets like those which exist in other areas. There is of course a good reason for this, but it does hamper development. If scientists can evaluate and compare their methods on the same data, this has a completely different weight than if everyone reports great results on their own, unpublished data. Comparability is thus lost. Not only the selection of the training data but also a small variation of the problem itself can have a great influence on the result. The MIMIC-Datensatz (Medical Information Mart for Intensive Care) exists in the English-speaking world and can be accessed in accordance with strict security hurdles, but unfortunately there is no equivalent in the German-speaking world. However, this would be extremely important in order to further promote development in our linguistic area.
The pooling of data would then even go one step further. I think there are already various efforts to pool resources, including research projects. In this respect, there is probably also a great deal of interest here at political level, and it will not work without political initiative.
The MACSS project has been used as a blueprint for your current BigMedilytics project, which is addressing increasing productivity in the healthcare system in 12 pilot projects. What is it all about in detail?
The BigMedilytics - Big Data for Medical Analytics project is a relatively large EU project, with Philips as consortium leader. It aims to significantly increase healthcare productivity by using Big Data technologies while reducing costs, improving patients' chances of recovery and increasing access to healthcare facilities. There are a total of 12 pilot projects, each dealing with different topics in this area. We are working on „Kidney Pilot“ together with Charité, the Hasso-Plattner-Institut (HPI), the AOK Nordost and Essen University Hospital. Our pilot is based on the BMWi-funded MACSS project. Within MACSS, we had three years to set up the platform. In the context of BigMedilytics, an evaluation can now be carried out. For example, we can investigate whether the platform can contribute to better treatment of patients. In other words alone, if we look at adherence to treatment. Apart from this, the DFKI and HPI are working on additional methods to analyse data, such as identifying high-risk patients, to help doctors do this.
In countries such as China, the use of pattern recognition, machine learning and AI is to some extent taking on forms which further fuel fears of technology in our society, although there are many counter-examples of use in everyone's interest. How do you see the potential here in Germany and in the European context?
When we talk about AI, I have the feeling that development is currently really fast. A great deal is happening in research, in a large number of areas. But industry is also driving the issue forward. Many large international companies are involved such as Google, Facebook, Uber but also Tencent or Baidu. Perhaps this is somewhat comparable to the Industrial Revolution at the beginning of the 19th century. There is a kind of gold-rush atmosphere and the belief in the potential of technology is enormous. Especially in China there seems to be a great euphoria, partly initiated by AlphaGo, the government, but probably also by the many people who are very open to a "new" technology.
When you talk about "fears of technology in our society", you are presumably thinking about possibilities of surveillance and social scoring in China. Yes, I suppose this potential is being exploited. Fortunately, things are different here in Germany. With our past, many people take a much more critical view of this and reject the "transparent person". On the other hand, the sharing and availability of information can also have a lot of benefits for us and our society - so let's ignore the surveillance aspect. Personally, I think there is a lot of potential here, both socially and economically. The question we have to ask ourselves is whether we want this. These are issues which all of us in society have to discuss, and data protection certainly plays an important role here. Maybe not everything that is technically possible should actually be implemented.
In this context, we will also for example be working on an exciting project as of the end of this year, in collaboration with the Friedrich-Alexander University of Erlangen-Nuremberg, the Leibniz University of Hannover and Charité, to investigate the ethical, social and legal aspects of artificial intelligence in medicine. While we in BigMedilytics are working on new prediction and risk models, their safe application and clinical usefulness must first be proven by studies. The question must be asked how patients can be protected from overtreatment if for example an algorithm is too sensitive.
Let’s take a look into the future: what will a hospital visit look like in 2050?
A hospital visit in 2050? As I mentioned earlier, developments in artificial intelligence are currently very rapid. On the other hand, mills in the medical field grind more slowly. New products first have to be certified and this can sometimes take a little longer, at least in the past. But perhaps this process will also change in the future. Who knows ...
Diseases will still exist in 2050, but I think certain processes will be more efficient. Due to the lack of skilled personnel more people will need to be treated with fewer personnel. In this respect, efficiency will certainly play a major role. Let’s hope that this will also apply to appointment management, in order to reduce waiting times at the doctor's, but also to a more efficient exchange of information between the doctor and the patient. Telemedicine will also be an important aspect in 2050 to make processes more efficient and to compensate for a shortage of skilled workers, for example in rural areas.
Furthermore, I am sure that AI methods have increasingly been finding their way into hospitals. This does not always have to be obvious, but at least many devices or programs will integrate appropriate methods, whether for the analysis of X-rays or to support the physician in the clinic. An AI will not replace a doctor so quickly, but small programs will hopefully help the physician find the best possible treatment for the patient.
Thank you very much for the interview