How Does The Tool Actually Work?

Healthcare Technical Breakdown ✨

Information extraction plays a crucial role in the clinical domain by enabling the automated extraction of relevant medical information from unstructured text sources such as electronic health records (EHRs), clinical notes, research articles, and drug labels. In this context, information extraction techniques facilitate the identification and extraction of key entities, relationships, and events pertaining to patient diagnoses, treatments, medications, adverse reactions, and other clinically relevant information. By automating the extraction process, healthcare providers can streamline clinical workflows, enhance decision-making processes, improve patient care outcomes, and facilitate medical research and analysis on a large scale.

These demos are trained to tackle the problem of information extraction in the medical field as a sequence labeling one, which is a fundamental task in natural language processing (NLP) that involves assigning labels to individual tokens in a sequence of text. To do this task, usually named entity recognition (NER), the BIO (Begin, Inside, Outside) schema has been used to annotate entities such as drugs and adverse effects within clinical text. Under this schema, each token in the input text is labeled with a tag indicating whether it represents the beginning (B), inside (I), or outside (O) of a named entity. This structured labeling scheme allows for the accurate identification and delineation of multi-token entities, which is essential for downstream tasks such as information retrieval, question answering, and clinical decision support systems.

How Does The Tool Actually Work?

LLMOps Pipeline ✨

The Large Language Model pipeline involves three steps as follows:

Data preprocessing
To process data from natural language to the BIO scheme we used a custom script in Python. This process takes each entity in the training datasets together with the character where they start and search in order of appearance in the text to assign the corresponding label in each case (Drug, Adverse Effect or Disease). Additionally, we implemented a simple sentence tokenization with the period (“.”) token.

Fine-tuning
The fine-tuning process for both tools was carried out in two steps, as follows: we fine-tuned one of the most powerful Large Language Models (BERT). A virtual machine with a GPU Tesla T4 with 27.3 gigabytes of available RAM was used during the training process.

Post-processing
As a post-processing we treated the subword tokenization inherent to the model used. Additionally, when the models extract two or more entities separately, but they are contiguous, we group them into a unique entity.