What is the appropriate logistic regression recall rate in the field of disease screening

To be honest, there is no one-size-fits-all standard answer to this question. The conventional floating range in the industry is between 85% and 98%. To what specific value it reaches, the core anchor point is always the "cost of missed diagnosis". Secondly, it must match the carrying capacity of the screening link, the cost of subsequent diagnosis and the public acceptance.

Don’t think that this range fluctuates widely. Anyone who has run a real-life project knows that the recall rate is never an indicator viewed in isolation. It is essentially a seesaw on the logistic regression classification threshold - if you lower the cutoff value for positive determination, the recall rate will naturally go up. It is nothing more than adding more false positive samples to see if you can withstand the subsequent pressure.

For example, we previously helped a prefecture-level city to develop a risk prediction model for HIV primary screening. Because missing a single case would not only cause the patient to miss the golden window for early intervention, but also cause subsequent transmission risks, the cost was too high, so we directly raised the recall rate to 97%. Even if the corresponding specificity dropped to 72%, it is completely acceptable - after all, people who are positive in the primary screening will undergo further nucleic acid confirmation, and a false positive will at most give the person a false alarm, which is nothing compared to the cost of missed testing.

But if it is a primary screening for prostate cancer for the general population, we will not put the recall rate so high. When we were working on a community screening project in a northern county, we calculated that if the recall rate reached 95%, nearly 30% of the people with false positives would need to undergo prostate biopsy. Not only does puncture carry the risk of infection and bleeding, but the pathology department resources in the county cannot handle so many samples. On the contrary, the really high-risk patients will not be queued up and the diagnosis will be delayed. So in the end, we increased the recall rate to 87%. On balance, the overall diagnosis efficiency is higher.

There is now a lot of controversy over this threshold in the industry. Take colorectal cancer community screening as an example. One group firmly believes that the recall rate should be raised to more than 95%. After all, the cure rate of colorectal cancer detected by early screening can reach 90%. It is a pity to miss the diagnosis. People with false positives can only have a painless colonoscopy, which is now very popular. , there is no pain; the other group feels that colonoscopy resources in the sinking market are too tight, and too many false positives will crowd out resources for people who really need it. It is better to reduce the recall rate to 90%, screen out the highest-risk groups first, and use resources wisely. Each side has its own implementation scenario, and no one can convince the other.

Our team encountered a big pitfall when we first entered the industry. At that time, we were obsessed with academic indicators and wanted to increase the recall rate of a diabetes preliminary screening model to 95%. As a result, community doctors complained in the first month after it was launched, saying that dozens of residents with false positives had to be notified to retest their blood sugar every day. People complained very much. Some even thought that the community hospital wanted to make money for examination fees, but was unwilling to cooperate with follow-up follow-up visits. Only later did I realize that when it comes to disease screening models, the more beautiful the indicators are, the more useful they are. The model must be based on the actual situation of the entire link, and the most appropriate one is the best.