ADRSpanishTool: A tool for extracting adverse drug reactions and indications with NLP

 

  • The question 
In the field of research, at the UPM and UC3M, I studied the relationship between drugs and adverse effects in texts.


  • The analysis

We propose an hybrid method to DDi detection that uses a machine learning approach based on support vector machines and a linguistic approach that combines a simplification method similar to that of Segura-Bedmar et al. (2011), a negation method similar to that of Chowdhury, MFM. and Lavelli, A., (2013) and rules based on the dependency tree provided by the Textalytics eHealth PoS. 

To obtain the relationships between drugs and their effects, we developed several web crawlers in order to gather sections describing drug indications and adverse drug reactions from drug package leaflets contained in the following websites: MedLinePlus, Prospectos.Net12 and Prospectos.org13. Once these sections were downloaded, their texts were processed using the Text-Alyticis tool to recognize drugs and their effects. 
As each section (describing drug indications or adverse drug effects) is linked to one drug, we decided
to consider the effects contained in the section as possible relationships with this drug. The type of relationship depends on the type of section: drug indication or adverse drug reaction. Thus for example, a pair (drug, effect) from a section describing drug indications is saved into the DrugEffect table as a drug indication relationship, while if the pair is obtained from a section describing adverse drug reactions, then it is saved as an adverse drug reaction. This database can be used to automatically identify drug indications and adverse drug reactions from texts. 

As gold standard we used the DDI corpus which is annotated with pharmacological substances as well as the interactions between them (Herrero-Zazo, M. et al., 2013; Segura-Bedmar, I. et al., 2013). This is the first corpus which includes pharmacodynamic (PD) and pharmacokinetic (PK) DDIs. For training, this corpus is composed of 572 articles collected from DrugBank, with 5675 sentences, and 142 PubMed abstracts, with 973 sentences, and together 4694 true drug interactions.  The number of positive instances are 12.98% of positives instances on MedLine and 14.57% on DrugBank. For test, this corpus is composed of 158 Drug-Bank, with 1301 sentences, and 33 PubMed abstracts, with 306 sentences, containing together 327 true drug interactions.

The parsing pipeline took advantage of the TextAlytics text processing API that provided syntactic and semantic information. In addition, a machine learning model using SVM was built using the java module jSRE.


  • The result
The final, published result was an ML model that could detect relationships between drugs and adverse effects.



Comments

Post a Comment

Popular posts from this blog

Building a Powerful Chatbot with OpenAI and Elasticsearch Integration

 In the realm of modern technological advancements, harnessing the capabilities of artificial intelligence (AI) has become indispensable. One such manifestation of AI prowess lies in the creation of intelligent chatbots that not only understand user queries but also interact with data repositories to fetch relevant information. In this article, we'll delve into the process of developing a sophisticated chatbot using OpenAI's cutting-edge technology and seamlessly integrating it with Elasticsearch for efficient data retrieval. Understanding the Components: Before diving into the intricacies of building our chatbot, let's dissect the key components involved: OpenAI: OpenAI provides state-of-the-art natural language processing (NLP) capabilities, empowering developers to build conversational AI agents. By leveraging OpenAI's API, we can imbue our chatbot with the ability to comprehend and respond to user queries effectively. Elasticsearch: Elasticsearch serves as a powerfu...
Read more

Legal Text Analysis Project: Leveraging NLP for Enhanced Understanding and Insight

Image
 A key point in the development of NLP solutions is the consideration of language proficiency. The terms and jargon used in specialized sectors, such as the medical sector, differ greatly from those used in the financial sector, which in turn are not comparable to those in the legal sector. The legal texts contain a series of specifications such as: The breadth of the domain in terms of textual typology. The variety of target groups. The linguistic features of the domain. Furthermore, in the latter case, not only is the legal terminology of the domain covered, but it also tends to co-occur with terminologies from all areas. The limited number of NLP resources and tools tailored to the overall domain. The predominance of English, since most of the resources and tools available are developed for processing texts in English. A slowed adoption of smart technologies in the legal and administrative sector compared to other sectors such as the biomedical or financial sector. The heterogen...
Read more

Streamlining Log Analysis with Logstash, Filebeat, and Elasticsearch

Image
 In the realm of IT infrastructure management, effectively monitoring and analyzing application logs is crucial for maintaining system reliability and performance. To address this imperative, I've devised a robust system harnessing the capabilities of Logstash, Filebeat, and Elasticsearch. At its core, this system is designed to seamlessly ingest, process, and index application logs into Elasticsearch for comprehensive analysis and visualization. Let's delve into the key components and functionality of this integrated solution: Log Ingestion with Filebeat: Filebeat serves as the lightweight shipper responsible for tailing application log files and forwarding them to Logstash for processing. Its efficient design ensures minimal resource overhead while guaranteeing real-time log collection from diverse sources. Data Processing with Logstash: Logstash acts as the central processing engine, facilitating data transformation, enrichment, and filtering before indexing it into Elastics...
Read more