The Hospital of Southern Norway is implementing artificial intelligence to read through medical records. The AI locates critical factors that doctors need to know about before operating on a patient. (02.07.2018)
The project was initiated a little more than two years ago, and is a collaboration between the hospital, the University of Agder, and Anzyz Technologies.
- The solution we are designing at the hospital is based on an algorithm we have developed at the university. This algorithm will be further developed and commercialized by Anzyz Technologies, says Ole-Christoffer Granmo, CTO of Anzyz and professor in artificial intelligence at the University of Agder.
The algorithm is designed to use artificial intelligence to understand the language used in medical records. These records contain all relevant medical information about the patient in question, including observations and comments noted by doctors. As with all medical research, they describe the patient in text-based language.
- If we can make artificial intelligence understand language, we have opened a whole new world of possibilities. And this is exactly what we have done. The self-learning algorithm we have developed can analyze how language is used in a huge amount of text, says Granmo.
Many possibilities in healthcare
Anzyz contacted the hospital about a potential collaboration because they saw huge potential in implementing artificial intelligence in interpreting language in the healthcare industry, and found a shared interest in using new technology in new ways to increase efficiency and accuracy. Researcher Geir Thore Berge and medical doctor Tor Oddbjørn Tveit at the hospital have been the ones using the Anzyz platform to create the artificial intelligence.
- They have enabled Anzyz to understand hospital jargon and the language used in medical records by feeding more than 800 000 medical records into Anzyz, says Granmo.
- Anzyz learns by looking at the context of words and expressions, and by doing this it can understand the meaning of words it does not know due to the context it is found in – just like people do. The purpose is to develop a support system for decision making that will make life easier for doctors, for example when a patient is scheduled for an operation. In this case doctors need a lot of information about the patient, such as if there are any allergies that can interfere with the procedure. Anzyz reads through and interprets all information about the patient from the records, and surveys critical information that the doctors should know about before the operation.
- Doctors often have to read through hundreds of pages before operating. You also have to be very thorough to make sure that you do not miss important information about a patient. This is of course very time-consuming, and accuracy often has to be sacrificed if time is of the essence. Anzyz can go through this process in mere seconds.
During a trial period of three months Anzyz has been used daily by 30 doctors and nurses at the hospital. This initial test shows that the system is very accurate: in 92 percent of the cases Anzyz found everything that was written about allergies.
- The project has proved to be very useful, and has therefore caused attention and interest among the other doctors at the hospital. We are therefore looking at other critical areas where Anzyz can be used, says Granmo.
If one has managed to solve the issue of medical language, one can link medical records with published literature and find the relevant medical research about a patient’s illness. New medical knowledge can also be discovered by combining bits and pieces of existing knowledge in new ways. Relevant information can be spread across many different scientific articles in very diverse fields, and it is impossible for a single person to locate all this information and put it together. Artificial intelligence enables us to do this.
- Say that you in one study discovered that Parkinson patients lack a certain chemical in the brain, and in another looked into the effect of a medicine against epilepsy that increase the production of this chemical. It would be extremely interesting to find out more about this connection, but to test the hypothesis from scratch would be both costly and time-consuming. However, if you link these two findings to medical records using artificial intelligence, you can easily locate patients that are suffering from both Parkinson and epilepsy, and have been given the medicine in question. It is then possible to test this kind of hypothesis in a few seconds, using already existing data. In this way what we are developing can contribute to revolutionize the way the healthcare industry works, says Granmo.
Read the article in Norwegian here.