NLP is Raising the Bar on Accurate Detection of Adverse Drug Events

NLP is Raising the Bar on Accurate Detection of Adverse Drug Events
 David Talby, CTO, John Snow Labs

Each year, Adverse Drug Events (ADE) account for nearly 700,000 emergency department visits and 100,000 hospitalizations in the US alone. Nearly 5 percent of hospitalized patients experience an ADE, making them one of the most common types of inpatient errors. What’s more, many of these instances are hard to discover because they are never reported. In fact, the median under-reporting rate in one meta-analysis of 37 studies was 94 percent. This is especially problematic given the negative consequences, which include significant pain, suffering, and premature death.

While healthcare providers and pharmaceutical companies conduct clinical trials to discover adverse reactions before selling their products, they are typically limited in numbers. This makes post-market drug safety monitoring essential to help discover ADE after the drugs are in use in medical settings. Fortunately, the advent of electronic health records (EHR) and natural language processing (NLP) solutions have made it possible to more effectively and accurately detect these prevalent adverse events, decreasing their likelihood and reducing their impact. 

Not only is this important for patient safety, but also from a business standpoint. Pharmaceutical companies are legally required to report adverse events – whether they find out about them from patient phone calls, social media, sales conversations with doctors, reports from hospitals, or any other channel. As you can imagine, this would be a very manual and tedious task without the computing power of NLP – and likely an unintentionally inaccurate one, too. 

The numbers reflect the importance of automated NLP technology, too: the global NLP in healthcare and life sciences market size is forecasted to grow from $1.5 billion in 2020 to $3.7 billion by 2025, more than doubling in the next five years. The adoption of prevalent cloud-based NLP solutions is a major growth factor here. In fact, 77 percent of respondents from a recent NLP survey indicated that they use ​at least one​ of the four major NLP cloud providers, Google is the most used. But, despite their popularity, respondents cited cost and accuracy as key challenges faced when using cloud-based solutions for NLP.

It goes without saying that accuracy is vital when it comes to matters as significant as predicting adverse reactions to medications, and data scientists agree. The same survey found that more than 40 percent of all respondents cited accuracy as the most important criteria they use to evaluate NLP solutions, and a quarter of respondents cited accuracy as the main criteria they used when evaluating NLP cloud services. Accuracy for domain-specific NLP problems (like healthcare) is a challenge for cloud providers, who only provide pre-trained models with limited training and tuning capabilities. This presents some big challenges for users for several reasons. 

Human language very contexts- and domain-specific, making it especially painful when a model is trained for general uses of words but does not understand how to recognize or disambiguate terms-of-art for a specific domain. In this case, speech-to-text services for video transcripts from a DevOps conference might identify the word “doctor” for the name “Docker,” which degrades the accuracy of the technology. Such errors may be acceptable when applying AI to marketing or online gaming, but not for detecting ADEs. 

In contrast, models have to be trained on medical terms and understand grammatical concepts, such as negation and conjunction. Take, for example, a patient saying, “I feel a bit drowsy with some blurred vision, but am having no gastric problems.” To be effective, models have to be able to relate the adverse events to the patient and specific medication that caused the aforementioned symptoms. This can be tricky because as the previous example sentence illustrates, the medication is not mentioned, so the model needs to correctly infer it from the paragraphs around it.

This gets even more complex, given the need for collecting ADE-related terms from various resources that are not composed in a structured manner. This could include a tweet, news story, transcripts or CRM notes of calls between a doctor and a pharmaceutical sales representative, or clinical trial reports. Mining large volumes of data from these sources have the power to expose serious or unknown consequences that can help detect these reactions. While there’s no one-size-fits-all solution for this, new enhancements in NLP capabilities are helping to improve this significantly. 

Advances in areas such as Named Entity Recognition (NER) and Classification, specifically, are making it easier to achieve more timely and accurate results. ADE NER models enable data scientists to extract ADE and drug entities from a given text, and ADE classifiers are trained to automatically decide if a given sentence is, in fact, a description of an ADE. The combination of NER and classifier and the availability of pre-trained clinical pipeline for ADE tasks in NLP libraries can save users from building such models and pipelines from scratch, and put them into production immediately. 

In some cases, the technology is pre-trained with tuned Clinical BioBERT embeddings, the most effective contextual language model in the clinical domain today. This makes these models more accurate than ever – improving on the latest state-of-the-art research results on standard benchmarks. ADE NER models can be trained on different embeddings, enabling users to customize the system based on the desired tradeoff between available compute power and accuracy. Solutions like this are now available in hundreds of pre-trained pipelines for multiple languages, enabling a global impact.

As we patiently await a vaccine for the deadly Coronavirus, there have been few times in history in which understanding drug reactions are more vital to global health than now. Using NLP to help monitor reactions to drug events is an effective way to identify and act on adverse reactions earlier, save healthcare organizations money, and ultimately make our healthcare system safer for patients and practitioners.

About David Talby

David Talby, Ph.D., MBA, is the CTO of John Snow Labs. He has spent his career making AI, big data, and data science solve real-world problems in healthcare, life science, and related fields. John Snow Labs is an award-winning AI and NLP company, accelerating progress in data science by providing state-of-the-art models, data, and platforms. Founded in 2015, it helps healthcare and life science companies build, deploy, and operate AI products and services.