What Claim Triage Is Predicting
A claims team may receive more files than adjusters can examine in depth each day. A prediction can help decide which files should be considered first, but only if the task is framed around the decision the team actually needs to make. The starting point is not a choice of algorithm. It is a clear statement of what one prediction describes, what outcome it concerns, and how the result fits the available review capacity.
For a claim triage task, the natural unit is usually one claim file at a defined point in its handling. A policy can generate multiple claims, and one claimant can be associated with multiple claims; neither fact makes the policy or claimant the right unit if adjusters make a separate review decision for each file. The prediction is an input to prioritization, not a decision that a claim is wrong or that a particular action will improve its outcome.
A binary outcome can be useful when the operational question is whether a claim meets a consistent definition of review need. For example, a team might define a positive outcome as “a specialist review is warranted under the standard review protocol within five business days of initial assignment.” The definition should distinguish need for review from review actually occurred: actual review can depend on past staffing and referral practices, while the intended triage question concerns which claims merit attention.
The prediction time matters because the team must be able to act on the prediction. If a score is to be generated when a claim is first assigned, its inputs must be information available then. A later note written during an investigation cannot be treated as if it were known at initial assignment. The task statement should make the prediction point explicit, even when the precise data design is developed separately.
Connect the Outcome to Adjuster Capacity
Let \(Y_i=1\) mean that claim \(i\) meets the agreed definition of review need, and let \(Y_i=0\) mean it does not. For information \(X_i\) available at the prediction time, a risk score can be written as:
Here, \(p_i\) is a predicted probability, not a certainty and not a finding that the claim is improper. If the team can review only \(k\) claims in the relevant period, it can rank eligible claims from highest to lowest \(p_i\) and send the top \(k\) for consideration. The period and the unit of capacity must match the operation: for example, a daily limit of claim files is not the same as a weekly limit of policies.
The ranking rule has a simple justification under specific assumptions. Suppose each selected claim uses one review slot, each review-positive claim has the same value to the team, there are no other constraints, and the probabilities are suitable for estimating expected outcomes. For a selected set \(S\), the expected number of review-positive claims is:
This follows from linearity of expectation; independence between claim outcomes is not required. If a selected claim has probability \(p_a\) and an unselected claim has probability \(p_b\), where \(p_b>p_a\), swapping them changes the expected number of positive claims by \(p_b-p_a>0\). Thus, under these assumptions, selecting the \(k\) highest probabilities maximizes the expected number of review-positive claims. If review costs, case values, or mandatory review rules differ, ranking by probability alone may not be appropriate.
Worked Example: Choosing Claims for a Two-File Review Queue
A hypothetical team predicts, at initial assignment, the probability that each claim will meet its standard definition of specialist-review need. The team can review two claims today. The predictions are: claim A, 0.72; claim B, 0.48; claim C, 0.31; and claim D, 0.64. Assume one slot per claim, equal value for each review-positive claim, and no other restrictions. Which claims should be prioritized, and how many review-positive claims should the team expect among them?
Method. Rank the claims by predicted probability, then select the top two because capacity is two. The ranking is A (0.72), D (0.64), B (0.48), C (0.31). The selected set is therefore A and D.
Expected count. Using the expected-count formula, the team expects \(0.72+0.64=1.36\) review-positive claims in the selected pair. This is an expectation, not a guarantee that exactly 1.36—or even exactly one or two—will be positive on a particular day.
Check. The next-best pair would replace D with B, giving \(0.72+0.48=1.20\), which is lower than 1.36. The selected pair therefore has the larger expected count under the stated assumptions. This method avoids filling a queue by file arrival order or by an arbitrary probability cutoff that ignores the two-slot limit.
Make the Target Match the Review Purpose
A target such as “was reviewed” may describe historical workflow rather than review need. If adjusters previously had limited capacity, some claims that warranted review may not have been reviewed. Conversely, a past referral can reflect a local practice that is no longer the intended standard. The task owner should decide what the model is meant to prioritize and define the outcome accordingly.
A useful target needs a consistent rule for deciding whether each claim is positive. The rule might refer to a documented protocol or a structured assessment by qualified reviewers. If the target cannot be determined consistently, a model score will not resolve that disagreement. The purpose here is not to decide whether a claim is valid or to estimate its dollar severity; it is to identify files that merit a specified form of review.
Worked Example: Separating Review Need From Past Referral
A hypothetical insurer wants to prioritize claims for a technical review. It examines six historical claim files. For each file, it records whether the claim was referred under the old workflow and whether an independent assessment applying the proposed review protocol says review was warranted.
| Claim | Referred under old workflow | Protocol says review warranted |
|---|---|---|
| 1 | Yes | Yes |
| 2 | Yes | No |
| 3 | No | Yes |
| 4 | No | No |
| 5 | Yes | Yes |
| 6 | No | Yes |
Step 1: Count each possible target. The old-workflow referral flag is positive for claims 1, 2, and 5, so its positive count is \(3\) out of \(6\). The protocol-based outcome is positive for claims 1, 3, 5, and 6, so its positive count is \(4\) out of \(6\).
Step 2: Compare the definitions. The two labels agree for claims 1, 4, and 5, or \(3\) of the \(6\) files. They differ for claims 2, 3, and 6. In particular, claim 3 and claim 6 would be treated as negative under the old referral label despite being positive under the proposed review-need definition.
Step 3: Choose the outcome that matches the task. If the intended queue is for claims that meet the proposed protocol, the protocol-based outcome is the relevant target. The old referral flag instead answers, “Was this claim referred under the old workflow?” That may be a valid question for another purpose, but it is not the same prediction task.
Check. The protocol-based positive fraction in these six files is \(4/6=0.6667\), or 66.67% when expressed as a percentage to two decimal places. This small historical summary does not establish future performance; it only makes the difference between the two candidate definitions explicit.
Work Within a Real Capacity Rule
Capacity should be represented as the constraint the team actually faces. If a team can review eight claims each day, its operational rule may be to rank that day’s eligible files and select up to eight. A fixed threshold such as “review every claim above 0.50” can select too few or too many files as claim volume and scores change. A capacity-based ranking is more directly tied to the queue, while a threshold can still be useful when it reflects a genuine minimum level of expected value or a required policy.
The capacity unit also affects what “top \(k\)” means. If each review slot is one claim file, a policy with two eligible claim files could occupy two slots. If the team can review at most one claim per policy, that is an additional constraint; simply taking the top \(k\) claim scores may then violate the rule. A clear task statement identifies whether capacity is measured in claim files, policies, or another operational unit.
Worked Example: A Daily Limit Is Not a Probability Cutoff
A hypothetical team receives ten eligible claim files in one morning and can review three. Their predicted probabilities of meeting the review-need definition, already sorted from highest to lowest, are 0.78, 0.61, 0.54, 0.47, 0.41, 0.36, 0.30, 0.24, 0.18, and 0.09. Assume each review slot has equal value and there are no other restrictions.
Step 1: Apply capacity. The team has three slots, so it selects the first three claims: 0.78, 0.61, and 0.54. The next claim, at 0.47, is not selected because capacity is full.
Step 2: Calculate expected positives. The expected number of review-positive claims is \(0.78+0.61+0.54=1.93\). The fourth claim would add an expected 0.47 positive claims if an additional slot were available.
Step 3: Check a common cutoff. A rule requiring a score of at least 0.50 would select only the first three claims here, but that agreement is incidental. If the top three scores had been 0.48, 0.42, and 0.35, a 0.50 cutoff would leave all three available slots unused even though each selected claim would have a higher predicted probability than the remaining claims.
Check. The top-three expected count is 1.93, and the team fills all three slots. This conclusion depends on the stated equal-value assumptions; if reviewing a claim has a cost that can exceed its benefit, a positive score by itself does not prove that the review is worthwhile.
Common Framing Errors
- Using a policy as the prediction unit when work is assigned by claim. The score may then fail to identify which particular claim file should be reviewed.
- Using “was reviewed” as a stand-in for “needed review.” The first can reflect past staffing and referral practices rather than the target the new queue is intended to serve.
- Confusing a prediction with an action recommendation. A high predicted probability identifies a claim as a priority under a stated rule; it does not by itself establish that review is mandatory or beneficial.
- Ignoring the capacity unit or time period. A score list cannot be translated into a queue until the team knows how many claim files, policies, or other units can be handled during the relevant period.
Check Your Understanding
For each scenario, identify the relevant task framing or apply the stated capacity rule.
- A team assigns a separate file to each claim and has one review slot. Two claims from the same policy have predicted review-need probabilities of 0.70 and 0.55; another policy’s claim has probability 0.62. Which claim should be selected if the slot is per claim and all claims have equal value? (Answer: Select the claim with probability 0.70.)
- A proposed target is “a claim was referred by an adjuster,” but the intended queue is for claims meeting a new technical-review protocol. What should define the target? (Answer: Whether the claim meets the stated technical-review protocol, not whether it was referred under the old workflow.)
- A team can review two claims. Their predicted probabilities are 0.44, 0.39, and 0.21, with equal value per positive review and no other restrictions. Which claims are selected, and what is the expected number of review-positive claims? (Answer: Select the 0.44 and 0.39 claims; expected count \(=0.44+0.39=0.83\).)
- A team has four review slots and a rule allowing at most one claim per policy. The highest-scoring claims are 0.83 for policy A, 0.76 for policy A, 0.71 for policy B, 0.68 for policy C, and 0.60 for policy D. Which four claims are prioritized? (Answer: Select policy A’s 0.83 claim, policy B’s 0.71 claim, policy C’s 0.68 claim, and policy D’s 0.60 claim; the second policy A claim is excluded by the one-per-policy rule.)
- A queue has three slots, and the top three claim probabilities are 0.49, 0.38, and 0.27. A manager proposes reviewing only claims with probability at least 0.50. Under equal value and no other restrictions, what is the capacity-based selection and its expected positive count? (Answer: Select all three claims; expected count \(=0.49+0.38+0.27=1.14\). The 0.50 cutoff would leave all three slots unused.)