How POSOS is using AI to improve the use of medicines by healthcare professionals and patients
In the age of Internet and rapid information access, it is not suspected that information retrieval on medicines could require sometimes hours, or even days, for healthcare professionals, such as doctors or pharmacists.
Summary
In the age of Internet and rapid information access, it is not suspected that information retrieval on medicines could require sometimes hours, or even days, for healthcare professionals, such as doctors or pharmacists.
It is, however, the daily reality of the 1,175,000 healthcare professionals working today in France, for whom information search can sometimes be a real obstacle. The information sources they have are difficult to exploit, although rich and numerous, which often hinders the filtering of relevant data. Facing difficulties of rapid access to reliable information, healthcare professionals rely more on their personal experience than on the data available in the scientific literature.
How to consequently help healthcare professionals to quickly have access to reliable and relevant information and help them in their approach?
This is exactly what our startup, POSOS, founded in 2017 in Amiens, proposes to solve by developing a technology, able to understand any question asked about a drug and automatically:
- identify the type of information sought and the context of the question,
- select from relevant information sources the relevant documents,
- Extract from these relevant documents the passages enabling to answer to the question.
The technology we are developing uses new methods of Natural Language Processing (NLP). Each of the steps described above is associated with an important NLP research area. The first step consisting in detecting the type of information sought is a classification and Named Entity Recognition (NER) problem. The second step consisting in selecting a list of relevant documents uses the Information Retrieval (IR) theory. Finally, the last step, enabling to find the relevant passages in these documents, is based on Machine Reading Comprehension (MRC). These different areas have all been revolutionized in recent years by the growing use of Deep Neural Networks (DNN).
Let us take, for example, the field of Information Retrieval. For a long time, conventional information retrieval systems consisted in mapping query keywords to corpus documents. Such systems, however, did not take into account any contextual information or semantic proximity between words. Thus, an effective biomedical information retrieval system should be able to link the term “antibiotic” with specific molecules such as “penicillin”, “cephalosporins” … Documents that frequently contain these names of molecules are thus certainly relevant for a request on antibiotics, although they do not explicitly mention the keyword “antibiotic”. It is in this respect that the application of neural networks has led to significant progress in the IR research area. Indeed, neural networks made it possible to obtain dense continuous vector representations of words (better known as word embeddings) by for instance being trained for each term to deduce the words surrounding it. The weights of the intermediate layers of such networks make it possible to obtain this continuous word representation. The underlying idea of this model is that two terms with a similar context will be geometrically close. Thanks to this word vectorial modeling, it is also possible to obtain representations of whole texts, enabling to simply compare geometrically the queries and documents of the corpus.
Nevertheless, NLP methods may have some limitations to obtain a general query system. These methods are often limited to a specific language and domain. We have therefore developed our own word embeddings in English and French using biomedical corpora and we are about to replicate our platform in other languages to become more competitive internationally. We are currently working on a neural optimization of our information retrieval system, to better take into account the words’ semantic information.
The tool we are developing at POSOS will benefit both healthcare professionals and patients. In the future, it will enable practitioner to have a reliable tool leading them to make enlightened decisions when prescribing a treatment. The patient can also be better served, thus reducing the risks associated with an inappropriate use of medicine, still causing nearly 200 000 deaths per year in Europe.
