What health tech gets wrong about clinical NLP – and how to get it right

Can’t listen to the full episode? Keep scrolling for some core insights. 

The problem with messy clinical data

“It’s shocking how non-standardized [clinical data] is… a single concept can be documented in a multitude of ways, depending on who’s doing the documentation. A simple concept like COVID-19, during the pandemic, was documented in 40 plus different ways… the variability leads to poor data quality.” – Haravu

Inconsistent clinical data isn’t just inconvenient – it has real-life impacts downstream. Take a basic hemoglobin A1C test. According to Haravu, physicians can document this test in various ways, such as:

• HbA1C
• HGB A1c
• Glycosylated hemoglobin
• HA1c

Now, multiply that variability across every provider, condition, and EHR system in the country, and you have a significant problem on your hands. That kind of inconsistency renders most clinical data unreliable at best, and wholly unusable at worst. That’s why you need clinical NLP – but not just any clinical NLP (more on that later).

80% of healthcare data is unstructured – that’s why clinical NLP matters

“As we know, we go to a doctor’s office, and the doctor is furiously typing on the keyboard… they’re typing a narrative about you… and each documentation style is different. Then you have to understand the context in which the medical concept is written.” – Haravu

Structured data, like lab results, matters, but most often, the richest information comes from unstructured documents like notes and histories – the information your clinician captures at the point of care. However, this kind of data is typically buried in the narrative and tricky to access and glean meaningful insights from. 

To make this information useful, clinical NLP must:

• Distinguish between assertions and negations (e.g. “no history of nausea” does not mean “has nausea,” even though the word “nausea” appears in the note)
• Understand the relationship between multiple diagnoses (e.g. diabetes that led to kidney disease that led to retinopathy)
• Interpret context across entire documents, not just isolated sentences.

This kind of data extraction requires NLP developed exclusively for healthcare needs with relevant context built in and routinely updated by clinical terminology professionals.

Behind the curtain: IMO Health’s approach to clinical NLP and AI

 “First, we focus on healthcare AI. We are not a generalist AI company. Two, we focus on a very specific discipline of AI called healthcare natural language processing… IMO Clinical AI is [a] platform for creating a natural language processing pipeline for largely clinical use cases.” – Haravu

Understanding IMO Clinical AI

IMO Clinical AI combines award-winning technology, a robust database of clinical terminology, and expert human oversight to enhance the accuracy of clinical insights and support data-driven healthcare decisions. It uses an advanced AI development platform, natural language processing (NLP) pipelines, and large language models (LLMs) in a platform designed specifically for healthcare settings to achieve results. 

IMO Health’s domain-specific clinical AI can:

• Identify documentation gaps in real-time at the point of care, leading to more accurate reimbursements for payers
• Help clinicians spot high-risk patients earlier, supporting more accurate risk assessments and enabling timely interventions that improve patient outcomes and reduce costs
• Provide reliable insights for health tech companies, supporting faster development workflows
• Support life sciences by extracting meaning from medical literature, real-world evidence, clinical notes, trial protocols, and patient data
• …and much more.

Perhaps most important, our experts continuously monitor IMO clinical AI-powered solutions to ensure the models don’t drift over time.

Foundational LLMs cannot handle complex clinical tasks

“If you try to code something to ICD 10, it’ll give you an answer – now you have to check whether it’s hallucinating or not.” – Haravu

General-purpose large language models (LLMs) are powerful, but they’re also risky in clinical settings due to a lack of domain training and a tendency to fabricate information (hallucinate).  These models are trained on a massive amount of data from the internet – but they’re only trained up to a certain point, meaning the data has an expiration date. Overall, these models are not adept at solving highly complex medical tasks that require deep semantic knowledge. 

IMO Health addresses these limitations by:

• Fortifying LLMs with proprietary, clinically curated terminology
• Proving critical context, including compound concepts with multiple code set mappings
• Leveraging advanced machine learning techniques, like retrieval-augmented generation (RAG)
• Continuously monitoring and updating information so it’s always current and usable

The result? Reliable NLP that can understand clinical intent, accelerate development processes, and effectively streamline manual tasks.

Why clinical NLP still requires human expertise

AI is a powerful tool – but it doesn’t replace clinical expertise.

Experts are necessary to:

• Monitor evolving diagnosis definitions
• Manage coding system updates
• Maintain editorial policies

At IMO Health, we employ numerous experts to support our unique healthcare NLP platform. Here’s a glimpse: