Author Archives: José Luis Martínez

About José Luis Martínez

Passionate about business around Natural Language processing application to solve real problems. Structuring unstructured data, even in big data environments. Partner at MeaningCloud.

Our experience on Adverse Drug Reactions (ADR) identification at TAC2017

MeaningCloud and LaBDA research group were present at the TAC 2017 conference held on November 13th – 14th at NIST headquarters in Washington. In the Text Analysis Conferences, research groups from all over the world were invited to develop software systems to tackle text analytics-related problems. This year, one task was devoted to the automatic identification of adverse drug reactions (ADRs) appearing in drug labels, including features defining the ADR, such as its severity or if it is characteristic of a drug class instead of just a given drug. There has been a specific subtask to link the identified ADRs with their corresponding MedDRA codes and lexical terms. More than 10 research teams have taken part in the project, all of them applying some kind of deep learning approach to the problem. Results show that it is possible to reach 85% accuracy when identifying ADRs.

We were delighted to present our text analytics-based system for ADRs identification on drug labels, which combines natural language processing and machine learning algorithms. The system has been built as a joint effort between MeaningCloud and LaBDA research group at the Universidad Carlos III de Madrid. Identifying ADRs is a basic task for pharmacovigilance, and that is the reason why the Federal Drug Administration (FDA) is involved in the funding and definition of the ADRs identification tasks in the framework of the Text Analysis Conferences. We have learned a lot these days (e.g., a BiLSTM deep neural network is the best choice for the purpose), and shared pleasant moments with our colleagues at Washington. We hope to be able to attend next year’s edition, which will focus on the extraction of drug-drug interactions (DDI), another interesting task aimed at detecting situations where the use of a combination of drugs may lead to an adverse effect.

We have now a dedicated business exclusively focused on the health and pharmaceutical sectors

Konplik.Health begins operations with the health-related assets from MeaningCloud, including its leading natural language processing, deep semantic analysis, AI platform, and adaptations specific to the life sciences.

Visit Konplik Health

Could Antidepressants Be the Cause of Birth Defects?

We agree that it is not typical at all for an Information Technology company to talk about antidepressants and pregnancy in its own blog. But here at MeaningCloud we have realized that health issues have a great impact on social networks, and the companies from that industry, including pharmas, should try to understand the conversation which arises around them. How? Through text analysis technology, as discussed below.

Looking at the data collected by our prototype for monitoring health issues in social media, we were surprised by the sudden increase in mentions of the term ‘pregnancy’ on July 10. In order to understand the reason of this fact, we analyzed the tweets related to pregnancy and childbearing. It turned out that the same day a piece of news on a study issued by the British Medical Journal about the harmful effects that antidepressants can have on the fetus had been published.
Continue reading →

Is Cognitive Computing too Cool to Be True?

According to IBM, “Cognitive Computing systems learn and interact naturally with people to extend what either humans or machines could do on their own. They help human experts make better decisions by penetrating the complexity of Big Data.” Dharmendra Modha, Manager of Cognitive Computing at IBM Research, talks about cognitive computing as an algorithm being able to solve a vast array of problems.

With this definition in mind, it seems that this algorithm requires a way to interact with humans in order to learn and to think as they do. Nice, great words! Anyway, it is the same well-known goal of Artificial Intelligence (AI), a more common name that almost everybody has heard about. Why change it? Ok, when a company is investing at least $1 billion in something, it must be cool and fancy enough to draw people’s attention, and AI is quite old-fashioned. Nevertheless, machines still cannot think! And I believe it will take some time.

How does Cognitive Computing work? According to the given definition, to enable the human-machine interaction, some kind of voice and image processing solutions must be integrated. I am not an expert on image processing, but voice recognition systems, dialog management models and Natuking-640388_1280ral Language Processing techniques have been studied for a while. Even Question Answering methods (i.e. the ability of a software system to return the exact answer to a question instead of a set of documents as traditional search engines do) have been deeply studied. We ourselves have been doing (and still do) research on this topic since 2007, which resulted in the development of virtual assistants, a combination of dialogue management and question answering techniques. Do you remember Ikea’s example called Anna? In spite of the fame she gained at that time, she is not working anymore. Perhaps, for users, that kind of interaction through a website was not effective enough. On the other hand, virtual assistants like Siri, supported by an enormous company as Apple, are gaining attention. There are other virtual assistants for environments different from iOS but they are far less known, perhaps because the companies behind them are quite smaller than Apple.

Several aspects of the thinking capabilities required by the mentioned algorithm have to do with the concept of Machine Learning. There are a lot of well-known algorithms which are able to generate models from a set of examples or even from raw data (in the case of unsupervised processes). This enables a machine to learn how to classify things or to group items together, like a baby piling up those coloured geometric pieces. So, combining Machine Learning and NLP models it is possible for a machine to understand a text. This process is what we call Structuring Unstructured Data (much less fancy than Cognitive Computing). That is, making your information actionable. We have been working on this during several years, but now it is called cognitive computing.

So, as you might imagine, Cognitive Computing techniques are not different from the ones we have already developed; a lot of researchers and companies have been combining them. And, if you think about it, does it really matter if a machine thinks or not? The relevant added value of this technology is helping humans to do their job with all the relevant information at hand, at the right moment, so they can make thoughtful and reasonable decisions. This is our goal at MeaningCloud.

Exploring Social Media for Healthcare Data

People enjoy sharing information through social media, including healthcare data. Yeah, it is true! And it constitutes the starting point of the research work titled ‘Exploring Spanish health social media for detecting drug effects’, which aims at following social media conversations to identify how people talk about their relation with drug consumption. This allows identifying possible adverse effects previously unknown related to these drugs. Although there is a protocol to communicate to the authorities the identification of a drug adverse effect, only a 5 – 20% of them are reported. Besides, conversations around drugs, symptoms, conditions and diseases can be analyzed to learn more about them. For example, it is possible to see how people search for specific drugs using social media, while others sell them, perhaps illegally. Many others talk about mixing alcohol with drugs or other illegal substances. Of course, one cannot believe everything that appears on the Internet —this is another issue—, but it can highlight some hypothesis for further research.

drugs

Some researchers from the Advanced Databases Group at Carlos III University of Madrid have carried out the mentioned study, designing hybrid models to capture the needed knowledge to identify adverse effects. The Natural Language Processing platform which supports the implementation of the analysis process based on such models is MeaningCloud. The customization capabilities provided by the platform have been decisive to include specific vocabulary and medical domain knowledge. As we know, the names of drugs and symptoms might be complex and, in some cases, difficult to write properly. The algorithm’s results are promising, with a 10% increase in recall when compared to other known algorithms. You can find further details in the scientific paper published by the BMC Medical Informatics and Decision Making Journal.

These developments have been part of the TrendMiner project, and are now available in the prototype website TrendMiner Health Analytics Dashboard, which shows people’s comments about antidepressants gathered from social media. The console displays the mentions of antidepressants and related symptoms and, by clicking on any of them, their evolution over time. Moreover, the source texts analyzed to compute those mentions are shown at the bottom, with labels highlighting the names of drugs, symptoms or diseases, and any relations among them. Such relations might say if a drug is indicated for a symptom or if a disease is an adverse effect of the mentioned drug. The prototype also allows searching by the ATC code (Anatomical Therapeutic Chemical Classification System) and the corresponding level according to this classification scheme. So, if you mark the ‘By Active Substance’ selector, you are searching any drug containing the active substance of the product you inserted in the search box. Furthermore, the predictive search functionality makes easier to find the right expression for a drug or disease.

Health and pharma companies can exploit their unstructured information

There are new kinds of data that are specific to the healthcare and pharmaceutical industries (such as electronic health records) as well as data science tools that allow us to extract valuable knowledge from that data.

 

With MeaningCloud, it is possible to identify the costs of medical treatments, their efficiency (cost, benefits, and risks), references to drugs, side effects, or long-term results. That is why our text analytics solution for the healthcare and pharma domains has so much potential.


Try our free demo!