Electronic health records (EHRs) are a treasure trove of information detailing oncology patients’ management and outcomes, but unlocking meaningful insights can be challenging. Structured data such as ICD10 codes and medication dispenses are useful but often lack the context crucial for a deeper understanding. Unstructured data, including physician notes, radiology and pathology reports, are rich with insights but must be laboriously reviewed, and at scale, this becomes infeasible.

Here we describe the development, validation, and scaling of automated Natural Language Processing (NLP) and Large Language Model (LLM) pipelines and go on to show real‑world applications of our ‘AI‑enhanced’ data model, including matching patients to clinical trial criteria with high accuracy and defining therapy‑related adverse events.

Using John Snow Labs’ NLP frameworks, we developed a comprehensive pipeline that performed optical character recognition, entity extraction, assertion detection, and relationship mapping from physician notes, radiology, pathology, and genetic reports focusing on temporal, anatomical, oncological, pathological, and treatment‑related entities and connections. Our dataset contained over 1.4 million physician notes and approximately 1 million PDF reports and scans. The technical implementation leveraged both classical LSTM‑based models and transformer‑based approaches (BioBERT).

Following deployment on a highly scalable Spark pool environment, we identified 113.6 million entities (averaging 80 entities per note) and 29.2 million relationships across 25 distinct relationship types (averaging 22 relationships per note). Comparison with a large manually reviewed truth set achieved high accuracy, with an entity extraction combined F1 score of ~93% and a relationship extraction F1 score of 88%.

By integrating NLP-derived elements into our standard data model, we can demonstrate an enhanced ability to match oncology patients to complex clinical trial enrollment criteria using a custom web application. Our matching performance was especially enhanced when cancer staging and biomarker findings were important inclusion criteria — as these are often missing from structured EMR data but can be recovered using NLP.

Although NLP greatly improves our data model, in certain contexts it performed poorly. We observed that the pretrained BioBERT model identified blood clots (thrombosis) with a precision < 0.5 — likely because of the many ways in which thrombosis can be described in notes and the lack of a causal drug relationship. As such, we developed a more sophisticated approach. Starting with all NLP‑predicted adverse events and flanking text, we used a locally cached base large language model, llama3 (LLM), and prompt engineering to classify text excerpts. The NLP/LLM hybrid approach performed significantly better (precision = 0.87) and allowed us to efficiently parse > 18k clinic notes in two hours.

These AI‑enhanced pipelines demonstrate the power of combining NLP and LLMs to extract critical information from EHRs, thereby enabling more efficient and accurate oncology research and clinical trial matching.