Your AI model predicts risk, but who decides when to act?

Meta description: A reliable probability becomes action only after you set a decision threshold; this explainer shows how cost changes model screening choices.

Tags: Model Evaluation, Decision Threshold, Calibration, Net Benefit

Featured image: abstract friendly illustration of a probability curve crossing a threshold line, no real people, no logos, no text.

The decision threshold is the moment a model stops being a ranking tool and starts deciding who gets acted upon. That step matters in any system that turns probability into action, including fraud scoring, demand forecasting, and anomaly detection. In a synthetic screening example, the model is built from 4,000 synthetic patients, of whom 448 are truly sick, or about 11 percent. It ranks sick patients ahead of healthy ones with an area under the receiver operating characteristic curve score of 0.84, and its probabilities are calibrated: when it assigns about 30 percent risk to a group, close to 30 percent of that group are sick. The model has already passed the earlier checks in this series: whether it can make a prediction, whether it can rank risk, and whether its confidence can be trusted.

A good probability is not a decision

To turn probabilities into yes or no calls, the default approach was used: every patient with predicted risk above 0.5 was treated as positive. At that cutoff, the model caught only a quarter of the disease, or 25 percent. Of the 448 truly sick patients, it flagged 110 and missed 338, while raising 56 false alarms.

The model itself did not fail. Its ranking remained intact, and its probabilities remained honest. What failed was the decision rule: a cutoff chosen without weighing consequences turned good risk estimates into a poor action. The patients were missed because 0.5 was the wrong place to act, and nobody had chosen it on purpose.

The default cutoff hides a value judgment

A threshold of 0.5 feels neutral, but it is not. It quietly says that a missed case and a false alarm cost the same. A screening model can make two mistakes: it can wave a sick patient through, a false negative, or flag a healthy patient for no reason, a false positive. Acting only when risk exceeds 50 percent is appropriate only if one missed case is worth exactly one false alarm.

Real decisions can involve more than that balance: the cost of running a test, the harm of treatment, or a person’s preferences. But the tradeoff between missed cases and false alarms is the right place to begin.

In many screening settings, the two errors are not equal. Missing an early cancer is not the same as calling one more patient back for a scan that turns out clear. When disease is uncommon, a 0.5 cutoff can be especially punishing. Low prevalence does not guarantee this outcome, but when a model gives only moderate separation, even a calibrated model can leave many genuinely sick patients below 50 percent. A rule that acts only above 50 percent then flags almost none of them.

That does not contradict the model’s ranking. The area under the receiver operating characteristic curve measures order: sick patients scoring higher than healthy patients. It does not say where those scores land on the scale from 0 to 1. A sharper model might push many sick patients above 50 percent; this one leaves 338 of the 448 below it, ranked above healthy patients but still short of the halfway mark.

A statistical cutoff is only a starting point

The classic data-driven way to pick a cutoff is the Youden index, proposed in 1950. It uses two rates: sensitivity, the share of truly sick patients caught, and specificity, the share of truly healthy patients correctly left alone. It chooses the threshold that maximizes J = sensitivity + specificity − 1, which is the same as maximizing their sum because −1 is constant.

On this model, the Youden index lands at 0.12. The picture changes completely. Sensitivity jumps from 25 percent to 77 percent, specificity is 76 percent, and the model catches 344 of the 448 cases instead of 110. But false alarms climb from 56 to 848.

The Youden index quietly assumes that one point of sensitivity and one point of specificity are worth the same. It is a statistical balance, blind to how common the disease is and what each error actually costs. It can be a reasonable reference point when no cost information exists, and the wrong anchor the moment cost information exists, which is almost always.

Cost ratio sets the threshold

The idea that reorganizes everything is simple: the curve of possible thresholds shows what can be achieved, but it cannot by itself say which tradeoff is acceptable. The missing question is how many times worse a missed case is than a false alarm.

Call that number R, the cost ratio. For a calibrated risk model, when the only things on the scale are the cost of a false positive and the cost of a false negative, the threshold that minimizes expected cost is:

act if predicted risk ≥ p_t, where p_t = 1 / (1 + R)

Here p_t is the probability threshold. If the two errors are equal, R = 1, then p_t = 1 / (1 + 1) = 0.50. That is exactly why the naive default catches so little. If a missed case is 9 times worse, R = 9, then p_t = 1 / (1 + 9) = 0.10. You act at just 10 percent risk, accepting more false alarms to avoid a miss. If a false alarm is feared more, say it is 4 times worse than a miss, so R = 1/4, then p_t = 1 / (1 + 0.25) = 0.80. Now you act only when you are 80 percent sure.

R is not a promise about headcount. It sets the relative cost of two errors, not the number of false alarms traded for each extra case caught. That depends on the model and how common the disease is. The number 9 is a statement of values, not a body count. The formula also assumes a plain yes or no decision. Capacity limits, patient-specific costs, or actions beyond act and do not act would need more than this one line.

The same model, same probabilities, and same patients can produce very different outcomes when only the cost statement changes. The threshold is where the human values enter, not the algorithm.

Does acting beat the obvious alternatives?

After choosing a threshold, one question remains: does using the model actually beat the obvious alternatives of referring everyone or referring no one?

This is what Decision Curve Analysis answers, a method from Vickers and Elkin in 2006. It rests on a reframing: the threshold you would act at is itself a statement of how you weigh the two errors. A clinician who would act at 10 percent is saying a missed case is worth about nine false alarms, so the harms should be weighed in that proportion.

That gives a single number called net benefit at any threshold:

net benefit = true positives / N − false positives / N × p_t / (1 − p_t)

Here N is the total number of patients evaluated. True positives are the good the model does; false positives are penalized in proportion to the same tradeoff. Computing this across thresholds and plotting the model, refer everyone, and refer no one shows whether acting on predictions offers a better balance than the alternatives.

Across thresholds roughly from 5 percent to 30 percent, the model delivers at least as much estimated net benefit as either alternative. At very low thresholds, referring everyone can make sense because the priority is to miss as few cases as possible. As the threshold rises, each unnecessary referral carries more weight, and eventually referring everyone produces less net benefit than referring no one. Referring no one always scores exactly zero in Decision Curve Analysis, a mathematical reference point, not a claim that missing every case does no harm.

In this synthetic experiment, the model performs at least as well as both alternatives across that relevant range. Only at the very lowest thresholds, below about 2 percent, do referring everyone and acting on almost any suspicion amount to the same thing. But estimated is the important word: the results describe potential value under the experiment’s assumptions. They do not establish that patients would actually benefit in practice. Showing that would take further validation and, ultimately, a prospective clinical study.

When to use each cutoff

If the costs are genuinely unknown, use the Youden index as a statistical reference point, not a justified default. Use it to explore tradeoffs while working with domain experts to set an appropriate decision policy.

If one error is known to be worse than the other, state that cost ratio and let it set the threshold. If the goal is a simple yes or no action from a calibrated probability, p_t = 1 / (1 + R) turns values into a rule. If the decision involves capacity, varied patient costs, or multiple possible actions, the single formula is not enough.

For readers, the practical meaning is that model evaluation does not end with a score. A trustworthy probability still needs a deliberate cutoff. The model can say how likely something is; only the decision-maker can say when that likelihood is worth acting on.