Failing FDA drug inspections are recorded later and less often than passing ones, biasing models trained and monitored on them
This paper demonstrates that FDA drug inspection classifications suffer from severe reporting bias, where failing inspections are recorded less frequently and with significant delays compared to passing ones, leading to substantial performance degradation in machine learning models and monitoring metrics that can only be partially mitigated by specific statistical adjustments or resolved by publishing the posting date.
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
Every year, the United States Food and Drug Administration sends teams of inspectors to pharmaceutical factories to ensure they are making medicine safely. These inspections are a critical part of public health, and the agency publishes the results. When an inspector finds no problems, the facility receives a passing grade. When serious issues are found, the facility receives a failing grade, which can lead to fines, seizures of products, or other legal actions. For anyone trying to understand how well these factories are doing, these public records are a treasure trove of information. They are dated, they cover hundreds of thousands of facilities, and they are free for anyone to use. This makes them perfect for training computer programs, known as machine learning models, to predict which factories might be at risk in the future. The hope is that by studying past records, these programs can spot trouble before it happens, helping regulators focus their limited resources where they are needed most.
However, there is a hidden flaw in how these records are kept that has gone unnoticed until now. The flaw is not in the inspection itself, but in how the final result gets written down and made public. When a factory passes an inspection, the agency records the result quickly and reliably. But when a factory fails, the process is much slower, and often, the result is never recorded at all in the public database. This creates a distorted view of reality for anyone using the data. A computer program looking at these records sees a world where almost everything is fine, because the bad news is either missing or arrives so late that the program has already moved on. This is not a small error; it is a fundamental bias that changes what the data actually says.
A researcher named Phani Kumar Balagam investigated this issue by looking at over ten thousand drug inspections that took place between 2018 and 2023. The study focused on a specific group of inspections that had been given enough time to be fully processed. The findings were stark. For inspections that ended with a passing grade, about seventy-four percent had a final decision recorded in the database. For inspections that ended with a failing grade, only about twenty-six percent had a record. This means a passing inspection is nearly three times more likely to appear in the public record than a failing one. Furthermore, when a failing inspection does get recorded, it takes a long time. On average, a passing inspection is logged in about eighty-six days, while a failing one takes over three hundred days. In many cases, the failing result never appears at all.
This delay and absence create a serious problem for the computer models that rely on this data. These models are often trained to look at a factory's history to predict its future. If a factory failed an inspection five years ago, a model should ideally know that. But because the record of that failure might not exist yet, or might not exist at all, the model cannot see it. The researcher rebuilt a prediction model to test what happens when the computer is forced to wait for the record to actually appear before using it. The result was a significant drop in the model's ability to distinguish between good and bad factories. The model's accuracy fell by a measurable amount, proving that the previous high scores were partly an illusion created by data that the model could not have actually seen at the time. The model had been using information that was not yet available in the real world.
The problem is even worse when trying to monitor these models after they are put to work. Regulators often check if a model is still working correctly by comparing what the model predicts against what actually happens. They look at the ratio of failures the model predicted versus the failures that actually occurred. Because the public records are missing so many failures, this check gives a wildly wrong answer. The researcher found that the standard check only sees about six percent of the actual failures that the model should have caught. It is like trying to measure the depth of a river by looking at only the ripples on the surface while ignoring the deep currents below.
The study also tested the standard mathematical tricks that statisticians use to fix missing data. One common method tries to guess how many failures are missing based on the information available. The researcher found that this method fails completely in this case. The missing failures are not just delayed; they are often gone forever. Because the data does not tell us why a record is missing, the computer cannot guess the truth. Another method, which involves waiting longer to see if the data arrives, helps a little but not enough. Even after waiting two years, the check still misses the vast majority of the failures. The study concludes that there is no way to fix this monitoring problem using the current public data. The missing information is simply not there to be found.
The researcher suggests that the solution is surprisingly simple. The FDA already has the information it needs; it just needs to publish one extra piece of data: the exact date when a classification decision was made visible to the public. If the agency added a simple field showing when a record was posted, it would allow computer programs to know exactly what they could and could not see at any given moment. This would stop the models from using future information to make past predictions and would allow monitors to understand exactly how much of the picture they are missing. Without this change, the public records will continue to show a world where pharmaceutical factories are safer than they actually are, and the tools designed to protect us will remain blind to the very dangers they are meant to find.
Drowning in papers in your field?
Get daily digests of the most novel papers matching your research keywords — with technical summaries, in your language.