How Regulatory-Grade Oncology AI Is Transforming Cancer Care

Forbes

4 days ago

David Talby, PhD, MBA, CTO at John Snow Labs. Solving real-world problems in healthcare, life sciences and related fields with AI and NLP.
For decades, the oncology field has faced an unfortunate truth: Extracting high-quality, structured information from clinical charts is a tedious, labor-intensive and largely manual task. Even as AI models have advanced, their outputs remain incomplete without human intervention. And what many people don't know is that behind every patient is a cancer registry specialist (CRS) spending hours reading through charts, identifying events, interpreting dates and ensuring accuracy for each case.
But as we approach regulatory-grade accuracy—a level of performance long considered the exclusive domain of highly trained human experts—that's all about to change. In the world of cancer data extraction, this means AI is hitting a consistent threshold of 95% accuracy. That figure isn't arbitrary; it's the benchmark achieved by experienced teams working meticulously, often with multiple levels of quality control.
Thanks to the combined power of healthcare-specific natural language processing (NLP) and large language models (LLMs), and a careful approach to model selection and orchestration, we're crossing that threshold in some of the most critical areas of oncology information, including tumor staging, grading and beyond. Here's why it matters.
Hidden Complexities Of Oncology Data
To appreciate the significance of this leap, it's important to understand the scale and complexity of the problem. A single cancer diagnosis involves hundreds of discrete data points: dates of imaging, biopsies, surgeries, therapies, pathology reviews and more. There are often dozens of potential diagnosis dates, and a specific rule determines which one is considered official for registry purposes. Even determining the primary cancer site or tumor grade can involve navigating contradictory information scattered across different documents.
Currently, filling out a registry case takes a herculean amount of time and effort. Registrars estimated taking approximately one hour and 15 minutes to complete an abstract for a simpler case and about two and a half hours to complete an abstract for a more complex case. This is done once a year for each patient, and with growing backlogs, data is often outdated by the time it's available for clinical decisions or research. The delay isn't just inconvenient. It's a barrier to real-time care optimization and scientific discovery.
Over the years, AI models have grown steadily more accurate. Best-in-class systems could extract relevant information from charts, but not reliably enough to replace human interpretation. They were assistive tools that were helpful, but not trustworthy enough to operate independently in regulatory contexts.
Why General-Purpose LLMs Fall Short
Now, with AI systems achieving 95%-plus accuracy on key fields without manual oversight, AI can replicate, and, in some cases, outperform, the gold standard achieved by expert cancer registrars. But not all models are created equally. These AI-driven tools are built specifically to tackle the unique challenges of healthcare, and oncology in particular. Rather than relying on general-purpose AI like GPT-4, which often struggles with domain-specific details, these models are trained on medical texts and structured to understand the nuances of clinical language.
It's tempting to believe that large, general AI models can solve these problems with simple prompts like, "Extract cancer diagnosis and treatment." But in practice, they fall short. Too often, they miss subtle distinctions, hallucinate relationships between entities or misinterpret clinical negations. While useful as a starting point, they lack the precision, stability and regulatory readiness needed for real-world healthcare applications.
The Power Of Medical Language Models
Healthcare-specific language models aren't just a tech upgrade; they're a foundation for the next generation of cancer care. What was once buried in notes and PDFs is now accessible, providing real, actionable insights. What this looks like in practice is automated case findings, real-time reporting and monitoring integrated into existing clinical workflows.
Achieving higher accuracy in entity recognition, better handling of negation and superior ontology mapping, domain-specific models produce results that are reproducible and explainable, which are key for auditability and trust.
Here are several ways regulatory-grade oncology AI is being applied:
• Tumor Registry Automation: Cancer centers are required to maintain registries of patients, including data on diagnosis, staging and treatment. Oncology models can scan pathology reports, read and decode them automatically, drastically reducing the need for manual chart review.
• Clinical Trial Matching: Finding eligible patients for a trial targeting a very specific cancer can be like finding a needle in a haystack. AI models can sift through thousands of records, pulling out the relevant biomarker and tumor type to flag potential candidates in near real time.
• Quality Monitoring: AI can flag when recommended treatments are missing. For example, if a patient doesn't have a recorded therapy plan, the system can alert the quality improvement team to investigate further.
• Adverse Event Tracking: Side effects can be buried in progress notes. AI can extract and monitor such events over time, alerting clinicians when recurring toxicities could signal a need to adjust therapy.
• Outcomes Research: For research teams comparing outcomes, AI tools can provide the structured data needed to stratify patients and link treatment patterns to survival trends.
Despite the obvious benefits, AI isn't a fix-all for oncology tracking. In rare cancers, evolving treatment protocols or atypical patient presentations, human registrars can be better equipped to contextualize and accurately code information that lacks precedent in training data.
Regulatory compliance, ethical considerations and quality assurance also demand expert oversight, ensuring data integrity and alignment with evolving standards. So, for now, human expertise remains vital to the accuracy and reliability of cancer registries. The role will just evolve with the technology.
With regulatory-grade AI for oncology, structured cancer data will become as current and accessible as the clinical notes they come from. Instead of data entry, registrars can shift their focus to more meaningful work, like quality assurance. In turn, patients will benefit from faster research and more responsive care. We're nearing the point at which AI is no longer just supporting our work—it's starting to do the work itself, and do it at a level healthcare professionals can rely on. It's just going to take time.
Forbes Technology Council is an invitation-only community for world-class CIOs, CTOs and technology executives. Do I qualify?