Clinical data abstraction is the work of reading a patient's records and turning the facts written in them into structured data, such as a field for the stage of a cancer or a yes-or-no field for a past heart attack. It is now often done by language models, with people checking the results. The data is used for clinical trials, quality measures, disease registries and care plans.
Most of the important facts in a record are written as free text. A doctor's note says, in a sentence, that the tumour has spread to two lymph nodes. A scan report describes a mass of 2.3 cm. Software that reads only structured fields, such as codes and lab values, cannot use these sentences. Reading clinical text with software is called clinical NLP.
Until recently, people did this reading by hand. Trained abstractors, often nurses or medical coders, read through each patient's file and filled in forms. Hospitals in the United States must send such data to cancer registries and to quality programmes. A large share of this manual work has been done in data-abstraction centres in India. It is slow: one complex cancer file can take an hour or more.
Language models now do the first reading. A large language model (LLM) is a model trained on huge amounts of text to read and write it. The software gives the model a patient's notes and reports, with a list of questions, such as "What is the stage of the cancer?" or "Has the patient received immunotherapy?" The model answers each question and quotes the sentence in which it found the answer. An abstractor then checks each answer against the quote, instead of reading the whole file. The check matters, because a model can give an answer that is not in the record at all. This is called AI hallucination.
Imagine a cancer centre in Tennessee that runs a trial of a new lung cancer medicine. The trial needs patients with stage IV lung cancer, a particular gene mutation and no earlier immunotherapy. The software reads eight years of notes for 6,000 lung cancer patients. It finds 140 who appear to qualify, each with the quoted sentences as evidence. Two research nurses review those 140 in a week, confirm 95, and pass them to clinical trial patient recruitment. By hand, the same search would have taken months.
The extracted facts then join the structured records in the healthcare data platform. With identifying details removed, they can also go to a real-world evidence (RWE) platform. Health plans use the same method for risk adjustment, to find illnesses that a doctor wrote in a note but never coded on a claim.
Clinical data abstraction differs from medical coding. Coding turns one visit into billing codes for a claim. Abstraction pulls many detailed facts out of years of records, for research and care.