Automated triage and screening systems are almost always configured to flag more cases than necessary. The asymmetry is deliberate and follows from the very different costs of the two ways such a system can be wrong.
The two errors have unequal consequences
A false positive sends someone for a review that turns out to be unnecessary, costing time, money and anxiety.
A false negative sends someone home with a condition that goes untreated, and the harm can be permanent.
Because the costs differ by orders of magnitude, a system optimised for overall accuracy is optimised for the wrong thing.
The threshold is where the judgement lives
These models output a score, and a threshold converts it into a decision to flag or clear.
Moving the threshold trades one error against the other directly: lowering it catches more true cases and generates more false alarms.
That setting is a clinical and institutional decision rather than a technical one, and it should be made by the people who bear both costs rather than shipped as a default.
Headline accuracy conceals the trade
A single accuracy figure is dominated by the majority class, and in screening the majority is healthy.
A system that cleared everyone would report high accuracy while catching nothing, which is why sensitivity and specificity are reported separately.
Comparing two tools on accuracy alone therefore tells you very little about which would be safer in use.
Over-referral has its own limit
Tuning for sensitivity works only while the downstream service can absorb the additional referrals.
Beyond that point, waiting times lengthen for everyone including the genuine cases, and the system has converted a detection improvement into a delay.
Capacity is therefore part of the threshold decision, which means the correct setting differs between institutions running identical equipment on identical software.
Vendors cannot set it for them, and a default shipped from the development site is a guess about a service the vendor has never seen.
Alarm fatigue is the failure that follows
Staff exposed to a high volume of alerts that mostly resolve as nothing begin discounting them.
Once that happens the system's sensitivity is irrelevant, because the flag no longer changes behaviour, and the true positives are dismissed alongside the rest.
Systems that survive in practice therefore report a graded urgency rather than a binary flag, and are tuned against whether clinicians still act on the output rather than against the detection rate alone.