Zum Inhalt springen

Deep neural networks – applications in healthcare

Current developments in health care like increasing patient numbers, due to the population ageing, together with the technical progress and its resulting flood of data, present physicians and researchers with new challenges. Medical knowledge is growing at an unprecedented rate and is becoming outdated regularly within a short period of time. However, these challenges also represent new opportunities.

The application of AI methods can contribute to making this knowledge usable and to supporting the increasingly complex medical practice. Indeed, this insight is not new: Medicine was already named decades ago, as one of the first practical fields of application for AI.  Algorithms with the cryptic names like PUFF1 or CADUCEUS2 formed the basis for the first commercial AI products at the end of the 1980s, such as the diagnostic database „Diagnosis“ 3. Progress in deep learning (DL) has triggered numerous technological developments in medicine in recent years, and a number of companies and scientists have turned to the field of AI. Nevertheless, earlier AI technologies have already been established in clinical application in Germany for many years.  So-called expert systems and here in particular knowledge-based systems are an integral part of daily work in outpatient and inpatient care.  Such applications are used, for example, for safety in drug therapy, in order to avoid contraindications or interactions between different drugs, and in the field of correct diagnosis and treatment coding. However, knowledge about this is still not widespread, and the confidence of the population is correspondingly low: 61 percent of around 1,000 respondents in an online study would rely on a diagnosis made by a physician with computer support.  If a finding is exclusively from a computer – for example through an AI application – only 12 percent of the respondents would not be sceptical. Regardless of this tendency, it is very likely that AI technologies will become significantly more established in patient care in the future. However, there are still a number of hurdles to overcome, such as the remaining challenges in dealing with large, complex and unstructured data volumes. For the widespread dissemination of practical AI applications in Germany, intelligent data protection solutions must also be found and fears of contact on the part of medical specialists and patients must be reduced. Patient trust can only be established through transparency of used methods and extensive data security as well as privacy efforts.4

Data growth – one of the central challenges

Even though forecasts for estimating the future global volume of data are difficult and differ to some extent, the trend is clear: new or improved technologies will increase the volume of data exponentially 5,6. This also applies to the medical sector. Improved, higher-resolution imaging methods are contributing to this, as is the increasing documentation of health-relevant information by medical personnel and by the patients themselves. This so-called digital self-measurement (quantified self) is carried out, for example, with fitness trackers and smartwatches or with increasingly used sensors in everyday objects such as scales, toothbrushes or insulin pens. Exponential data growth is accompanied by a comparatively moderate increase in the number of scientific publications.  The number of published articles and journals has grown by an average of three percent per year over the past decades; a large part of these publications is produced in the medical field7.The reason for this is a general expansion of the scientific community and consequently an increasing number of scientists. It can be assumed that although a multiplication of the available data is to be expected, the knowledge actually derived from it in the form of publications will grow more moderately.  However, as in the past decades, the progressing networking and better exchange of scientific results will lead to more and more knowledge being individually available. This means, that individuals will hardly be in a position to act at the current level of knowledge at all times. Medical knowledge has already reached a level that makes it almost impossible for doctors and medical staff to keep up to date. It is difficult for physicians to know and apply all treatment strategies and their applications equally. In addition, with the rapid increase of medical knowledge, supposed certainties can become obsolete overnight. AI-supported programs, which always incorporate the latest findings by means of self-learning algorithms, could provide a remedy. Practice makes perfect: machine and in-depth learning. The immense progress made by AI in recent years is essentially based on a combination of supervised learning with the use of deep learning approaches. In this interaction, a training data set is used as a starting point for the optimization of an algorithm. The larger the underlying data set, the more precise the algorithm can work. AI methods are considered to have high hopes, especially in the area of analysis of large unstructured and fast moving data (big data).  AI can thus contribute to the investigation of large amounts of data with regard to statistical relationships and thus help to gain new scientific knowledge – for example for predicting the effects of therapy, as a clinical decision-making aid and in monitoring drug safety8. So far, however, the necessary bioinformatics evaluation and practical use of complex data is still in its infancy.  Although the human genome can be sequenced for less than 1,000 US dollars, only a fraction of the huge amounts of data generated can be correctly interpreted in connection with a disease and used for the diagnosis or treatment of patients in the sense of a personalised medicine. Using various DL approaches, the initial quality of the genome data collected (FDA 2016) and their interpretation will be improved9. One hurdle that still exists when linking large amounts of data, for example from patient files, is the poor quality of data collection and classification. So far, there is only exemplary evidence that the use of Big Data technologies in hospitals has a practical advantage. A regular application in clinical everyday life is still far away.  However, AI in connection with DL is not only discussed in Big Data analyses, but also in a multitude of other potential medical applications.


Radiology: AI in the four-eyes principle

In medical imaging, so-called expert systems have been used for several decades.  In English, this is often referred to as computer aided detection and computer aided diagnosis (CAD). The bibliographic reference database PubMed lists studies in this field beginning in the 1970s. These computer programs support radiologists in interpreting the image data.  Computer tomographic image data can consist of several thousands of individual images; therefore, the radiologist may need a lot of time to produce a finding. Since this also requires a very high concentration, careless mistakes can quickly creep into the work. CAD supports radiologists in these cases with pattern recognition to identify relevant individual images and to point out conspicuous features. It has proven practical to simulate the classic „four-eyes principle“ with CAD:  The radiologist first evaluates the images, and then the computer uses an algorithm to check which image sections should be viewed more closely 10. The biggest difference between this common practice and the current developments in the field of DL is that people programmed the algorithms on which such expert systems are based, so that they could only achieve a certain degree of complexity.  DL now makes it possible for the algorithm to automatically extract findings from each analyzed data set, which then flow into the analysis of the next data set.  In this way, the sensitivity and specificity of the results are continuously optimized.  However, these new developments primarily serve to increase efficiency and shorten the time required by doctors for diagnosis and therapy.

Radiology: Quantification of tumours

However, the application possibilities of AI in radiology are not limited to the marking of such abnormalities.  As soon as, for example, a tumour or lesion is identified in a patient, it is measured.  In addition to size and volume, consistency and structure are also determined.  Such a measurement, which is carried out manually by the radiologist, is time-consuming because, among other things, the tissue boundaries in each sectional image must be precisely determined so that the size can be calculated later. It is obvious that such activities can be automated with AI methods.

Radiology: Monitor and analyse development of disease

DL algorithms can also help to analyse the development of a disease. Thus, it is already possible today for programs from the electronic patient file to automatically call up the radiological images from the archive and at the same time the corresponding current sectional image, so that the radiologist can compare the images. The software can also easily scale and align these images.


It is already possible to scan for skin cancer using intelligent algorithms. These have been trained by thousands of data sets comprised of healthy and effected skin. Programs using these algorithms have been shown to predict the malignancy of a tumour much faster and more accurately than humans. Indeed, there is already a smartphone app released, which is capable of doing so. Yet, is only available to a selected audience of researchers.

Workflow optimization

KI can also be used to optimize the entire clinical workflow. An automated evaluation of the image data after anomalies allows patients with acute treatment needs, for example, to be identified more quickly and prioritized accordingly by the computer. In addition, it is conceivable that the AI algorithm could also examine the data for abnormalities which is not directly related to the patient’s complaints.

Electronic patient file

Some of the AI applications described using the example of radiology can only be realised in combination with an electronic patient file (EPF), for example the comparison of new radiological images with already existing ones.  This begins with the digitalization of previously paper-based documents, which are then evaluated and structured using algorithms for free text recognition.  In this way, information can be searched electronically in an EPF and quickly found. In addition, drug reactions and contraindications can be determined and identified in a targeted and comprehensive manner. KI can help to determine the best individual therapy on the basis of the data stored in an EPF.

In addition to classic AI applications such as drug therapy safety and diagnostic coding, machine and in-depth learning has already been used for some time in the processing of patient files. Not only hospitals use these AI methods, but also insurance companies (H20.ai 2017).  In Germany, however, applications of these AI techniques are currently an exception due to the often decentralized storage of patient data.

In recent years, developments around the Watson Health AI platform developed by IBM have made a name for themselves. According to manufacturer data, six countries used Watson in clinical care in 2017 (Bloomberg 2017).  Analyses of ten different types of cancer are now possible. And this can also be done online: On the homepage of an Indian hospital chain, Watson can obtain a kind of second opinion report on the optimal treatment regime (Manipal Hospitals o. J.) after uploading its own patient file.  In a model project, the hospital operator Rhön-Klinikum AG wanted to use Watson for text and document recognition in Germany as well, but discontinued this project in 2017 in order to continue it with another provider.

Early detection and prevention: prevention is better than cure

On the basis of rapidly progressing findings on the molecular mechanisms of a wide variety of diseases and their diagnosis with the aid of AI, the vision of recognising and treating diseases as they develop is approaching. In many indications, early therapy could offer the best chances of cure or even completely and permanently prevent the outbreak of the disease.  One example from today’s healthcare is the treatment of people with high cholesterol who do not have symptoms of disease. In many cases, cardiovascular diseases can be prevented with statins. However, cardiovascular problems are dependent on a large number of other influencing factors, and it is therefore not clear which of the people treated really benefit from the medication, which medication may even be more harmful and for which additional preventive measures are urgently required. In a study conducted by the University of Nottingham, four different AI systems were compared with the medical guidelines used to date in order to predict from a large clinical data set which individuals will suffer a cardiovascular event, such as a heart attack, in the next ten years. All four AI systems were superior to the guidelines.  The algorithm that was trained via neuronal networks performed best.  He not only correctly predicted 7.6 percent more disease events, but also triggered 1.6 percent fewer false alarms due to incorrect results.  In the total of 83,000 patient files examined, a further 355 persons could have been identified for whom preventive treatment or lifestyle changes could have prevented a cardiovascular event under certain circumstances. CI could thus save lives in this concrete application 11.

CI might even be able to predict other complex diseases with multifactorial triggers, such as neurodegenerative diseases.  In an Italian study, an AI could be trained in such a way that brain scans could be used with great reliability to determine whether a patient was likely to develop Alzheimer’s within a decade. Magnetic resonance imaging was used to detect the smallest changes in the connections between different brain regions.  Although Alzheimer’s dementia cannot be cured so far, a diagnosis in the symptom-free early stage would still have some advantages. It would, for example, enable the affected persons to change their lifestyle in order to reduce known risk factors for the disease. In addition, there are indications that treatment with medication available today is more effective the earlier it is used in the course of the disease.  Furthermore, the diagnosis of the first non-specific symptoms can be used to differentiate Alzheimer’s disease from other forms of dementia. In the long term, AI methods could help to prevent the development of diseases, which would be tantamount to a paradigm shift from the current reactive disease care to preventive health care. In order to train AI for this goal, ideally excellently structured data of very many people would have to be available over as long a period as possible – as will be the case, for example, in the „All of US“ cohort study of the US American NIH. The study is to provide voluntary support for one million or more people over many years and record their state of health, environment and lifestyle in detail (National Institutes of Health 2018).

Learn more here:


1 Aikins, J. S., Kunz, J. C., Shortliffe, E. H.; Fallat, R. J. (1983): PUFF: an expert system for interpretation of pulmonary function data. Computers and biomedical research, an international journal, 16 (3), 199–208.

2  Banks, G. (1986): Artificial intelligence in medical diagnosis: the INTERNIST/CADUCEUS approach. Critical reviews in medical informatics, 1 (1), 23–54

3 Ärzte-Verlag, D. (1989): Ein Programm zur Diagnosehilfe (47), 104. Zugriff am 21.02.2018. Verfügbar unter https://www.aerzteblatt.de/pdf.asp?id=10993

4 Wittpahl, V. (Ed.), (2019) Künstliche Intelligenz. Technologie, Springer Berlin Heidelberg, Berlin, Heidelberg.

5 Reinsel, D., Gantz, J.; Ryding, J. (IDC, Hrsg.) (2017): Data Age 2025. The Evolution of Data to Life-Critical. Don’t Focus on Big Data; Focus on the Data That’s Big. Zugriff am 08.02.2018.

6 Schlömer, L. (2016): Spezialisten für Data Science händeringend gesucht. Online verfügbar unter https://www.digitale-exzellenz.de/spezialisten-fur-data-science-handeringend-gesucht/, zuletzt geprüft am 22.06.201

7 Ware, M.; Mabe, M. (2015): The STM Report. An overview of scientific and scholarly journal publishing. Celebrating the 350th anniversary of journal publishing (4. Aufl.) (: Internatio-nal Association of Scientific, Technical and Medical Publishers, Hrsg.). The Hague. Zugriff am 05.01.2018. Online verfügbar unter https://digitalcommons.unl.edu/cgi/viewcontent.cgi?referer=https://www.google.de/&httpsredir=1&article=1008&context=scholcom, zuletzt geprüft am 22.06.2018.

8 Lee, C. H.; Yoon, H.-J. (2017): Medical big data. Promise and challenges. Kidney research and clinical practice, 36 (1), 3–11. https://doi.org/10.23876/j.krcp.2017.36.1.3

9 Beyer, D. (2016): Deep learning meets genome biology. Online verfügbar unter https://www.oreilly.com/ideas/deep-learning-meets-genome-biology, zuletzt geprüft am 22.06.2018

10 Castellino, R. A. (2005): Computer aided detection (CAD). An overview. Cancer imaging : the official publication of the International Cancer Imaging Society, 5, 17–19. https://doi.org/10.1102/1470-7330.2005.0018

11 Weing, S. F., Reps, J., Kai, J., Garibaldi, J. M.; Qureshi, N. (2017): Can machine-learning improve cardiovascular risk prediction using routine clinical data? PloS one, 12 (4), e0174944. https://doi.org/10.1371/journal.pone.0174944.

Sei der Erste der einen Kommentar abgibt

Schreibe einen Kommentar

Deine E-Mail-Adresse wird nicht veröffentlicht. Erforderliche Felder sind mit * markiert.