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

In fact, there is no one-size-fits-all standard value. The core anchor point is "how much is the cost of missed diagnosis." The recall rate of routine low-risk chronic disease initial screening can reach 85%, which is basically sufficient. If it involves screening for fatal diseases and severe infectious diseases, it must be at least 95% or above. In extreme scenarios, the accuracy rate may even be raised to 98% under the pressure of plummeting.

Logistic regression itself outputs probabilistic results. The recall rate and precision rate are like seesaws. If you lower the threshold for positive determination, the recall rate will naturally go up. It is nothing more than a bunch of false positives. How to choose depends entirely on whether the scenario can accept the cost. Two years ago, when I helped the community health service center to create a preliminary screening model for elderly hypertension, I initially used the default 0.5 as the threshold, and the recall rate was only 78%. When I showed it to a clinical doctor, I was immediately rejected. They said that the preliminary screening was to cast a wide net, missing an elderly person with uncontrolled hypertension, and it would be fatal in case of a stroke. Calling a few more false positives to come back for retesting was nothing more than spending more than ten minutes to measure blood pressure and draw blood, which is completely acceptable. Later, I adjusted the threshold to 0.32, and the recall rate directly increased to 92%, while the precision rate dropped by 14%. After it went online, the feedback from doctors was very good. Basically, all patients with latent hypertension who had been missed before were fished out.

However, it is not necessary to increase the recall rate in all scenarios. I know many friends who make consumer screening products and do not agree with this logic at all. They sell screening services to ordinary C-end users. If the recall rate is too high and there are too many false positives, users will be scared to death when they get a positive result. They will go to the hospital for a review and nothing will happen. They will then give you a one-star complaint and say you are causing anxiety. There used to be a team that did preliminary screening of fecal occult blood for intestinal cancer. In order to meet clinical requirements, the recall rate was set to 97% at the beginning. As a result, 8 out of 10 users who reported positive were retested. The number of complaints exploded in the first month after it was launched. Later, they had to adjust the threshold back and the recall rate was reduced to 90%. False positives were directly cut in half. On the contrary, user reputation and clinical acceptance increased.

Another point that is easily overlooked is that you should not lose the interpretability of logistic regression just to pile up the recall rate. After all, in disease screening scenarios, what clinicians most appreciate is that logistic regression can clearly explain "why this person was tested positive." If you try to increase the recall rate by 2% and stuff a bunch of features that have nothing to do with the logic of the disease, no one will dare to use it no matter how good the data is. I have seen someone make a diabetes screening model before. In order to make up for the indicators, the user's zodiac sign and mobile phone model were stuffed into the feature pool. The final recall rate was increased by 1.8%. However, the clinic directly rejected it, saying that it did not make sense at all. Who dares to give such a report to the user?

To put it bluntly, the appropriate recall rate is never determined by the algorithm team behind closed doors and brushing up indicators. It has to involve clinicians, operations, and even user representatives to sit down and calculate the cost of missing a patient, the cost of one more false positive, and the adjusted threshold after calculation is truly appropriate. Sticking to a certain industry standard value can easily lead to problems.