Auditing Data Leakage Candidate Coverage and Calibration in Clinical Abbreviation Disambiguation Benchmarks
This paper introduces an auditable evaluation protocol for clinical abbreviation disambiguation that exposes how high benchmark accuracies often mask data leakage and candidate coverage issues, demonstrating that reliable clinical NLP assessment requires separate reporting of leakage robustness, candidate support, and calibrated reliability rather than relying on a single accuracy score.
Original paper licensed under CC BY 4.0 (https://creativecommons.org/licenses/by/4.0/). This is an AI-generated explanation of the paper below. It is not written or endorsed by the authors. For technical accuracy, refer to the original paper. Read full disclaimer
In the vast, unstructured world of medical records, doctors and nurses write notes that are dense with shorthand. To save time, they use abbreviations, but these shortcuts are often ambiguous. A single short string of letters might mean one thing in a cardiology report and something entirely different in a neurology note. For computers to help doctors, they must first learn to solve this puzzle, a task known as abbreviation disambiguation. Researchers have built public tests, or benchmarks, to see how well computer programs can guess the correct meaning of these shortcuts. These tests usually produce a single number, a percentage, that is meant to represent how accurate the computer is. If a program scores ninety-five percent, it is often assumed to be nearly perfect. However, this number can be misleading if the test itself contains hidden flaws, such as repeating the same sentences in both the learning phase and the testing phase, or if the test removes the hardest cases before the computer ever sees them.
A team of researchers set out to audit these tests, not just to see if the computers were smart, but to see if the tests were fair. They focused on a widely used collection of medical notes containing seventy-five different abbreviations. Their goal was to build a new way of evaluating these systems that looked behind the curtain of the final score. They found that the standard way of running these tests often hides two major problems. First, the same sentence can accidentally appear in both the training data, where the computer learns, and the test data, where it is graded. This is like giving a student a practice exam that contains the exact same questions as the final test; the high score reflects memory, not understanding. Second, the tests often delete rare meanings of abbreviations because there are only a few examples of them. This creates a false sense of security, because in the real world, a computer will eventually encounter these rare cases and have no correct answer to choose from.
To fix this, the researchers designed a strict protocol that acts like a rigorous quality control check. Before splitting the data into learning and testing groups, they scanned for duplicate sentences and removed the extras, ensuring that every sentence in the test set was truly new to the computer. They also refused to delete the rare meanings. Instead, they moved these difficult cases into a separate "stress test" group. This group allowed them to see what happens when a computer is asked to guess a meaning that it has never been trained to recognize. They also checked the computer's confidence. A good system should not only guess correctly but also know when it is guessing. The researchers measured whether the computer could tell the difference between a situation where it had a good answer and one where it was forced to guess blindly.
When they applied this new, stricter protocol to the seventy-five abbreviations, the results were revealing. The computer systems still performed well, but the researchers discovered that the high scores reported in the past were not entirely due to the systems' intelligence. By removing the duplicate sentences that had leaked between the training and testing sets, the scores dropped only slightly, by about two hundredths of a percentage point. This small change suggested that while the data leakage was present, it was not the main driver of the high scores in this specific dataset. However, the real story emerged when they looked at the rare meanings. When the computer was forced to choose from a list of options that did not include the correct answer, it still confidently guessed wrong more than half the time. Specifically, when the correct meaning was missing from its list of choices, the system still passed a confidence check designed to filter out bad guesses in fifty-eight percent of those cases. This means the computer was often very sure of its wrong answers.
The study also compared different types of computer models, looking not just at accuracy but at how much computing power they required. They found that a more complex model did not necessarily perform better than a simpler one, and when it did, the improvement was so small that it might not be worth the massive increase in time and energy needed to run it. One model was forty-two times larger and hundreds of times slower than another, yet it only offered a tiny edge in performance. This highlights that a single number like "accuracy" is not enough to judge a system. It hides the cost of running the system and fails to tell us if the system is reliable when it faces the unknown.
The researchers concluded that the way we evaluate medical artificial intelligence needs to change. We cannot rely on a single score to tell us if a system is ready for the real world. Instead, we need a package of evidence that includes how the test was built, whether the system can handle rare cases, and how much it costs to run. By separating these different factors, doctors and developers can make better decisions. They can see if a system is truly robust or if it is just good at memorizing the test. In the end, the goal is not just to build a computer that gets a high grade on a test, but to build a tool that can be trusted when a doctor is making a critical decision based on a confusing note. The researchers showed that by auditing the test itself, we can stop trusting numbers that look good but might be hiding the truth.
Drowning in papers in your field?
Get daily digests of the most novel papers matching your research keywords — with technical summaries, in your language.