Inconclusive Diagnostic Results: Keep the Middle Row
October 7, 2026 · 7 min read · Aqrab editorial team
A paper reports sensitivity and specificity for a test that, in practice, returns high, intermediate, or low results. Where did the middle group go? If the analysis excludes it or calls it “negative” without saying so, the headline accuracy may describe a different test from the one clinicians actually use.
The 2×2 Table Is Not Always Enough
Consider a deliberately hypothetical test evaluated in 100 people with the target condition and 100 without it. These are conditional counts chosen for teaching, not prevalence or performance data for a real test.
| Valid result | Among 100 with condition | Among 100 without | Result-specific LR |
|---|---|---|---|
| High | 60 | 15 | 4.00 |
| Intermediate | 20 | 20 | 1.00 |
| Low | 20 | 65 | 0.31 |
If a patient’s pretest probability were 20%, their pretest odds would be 1:4. After a high result, multiplying the odds by 4 gives a post-test probability of 50%. An intermediate result with LR 1 leaves it at 20%. A low result with LR 20/65 gives about 7.1%. These are arithmetic scenarios, not treatment or diagnostic recommendations.
The practical takeaway
Ask for a count of every result category by reference-standard status. A valid middle result may be informative, uninformative, or something in between; do not assume its LR is 1 without data.
Intermediate Is Not the Same as Failed
A valid intermediate result falls on the test’s scale and can have its own result-specific likelihood ratio. An uninterpretable result—such as one caused by an inadequate specimen or technical failure—is a different problem. Its frequency and causes should be reported separately, with the consequences for the intended workflow made clear. Neither type should disappear from the participant flow just because a binary accuracy statistic is easier to calculate.
Moving intermediate results into “positive” or “negative” changes the threshold and the question. Excluding them can make reported performance look better if the remaining cases are easier to classify. The direction and size of any change depend on the data; the table above does not quantify bias in a real study.
Five Questions for a Diagnostic Paper
- Result scale: What raw or ordered results did the test actually return before categories were combined?
- Disposition: How many results were intermediate, indeterminate, uninterpretable, or missing, and why?
- Reference standard: What was the target-condition status for each category, and how often was that status unavailable?
- Analysis: Were uncertain results excluded, grouped, repeated, or assigned to one side of a cutoff? Was this prespecified?
- Use case: What happens to a patient with a middle or failed result in the actual care pathway?
Reporting a full 3×2 table for three valid result bands can preserve information a 2×2 summary hides. Technical failures still need their own denominator and accounting; a result-specific LR cannot be assumed for a result that was never produced.
Where Aqrab Fits
When reviewing a diagnostic study or clinical-AI validation report, use Aqrab to pressure-test how the authors handled middle and failed results. A methods critique can identify missing denominators and unsupported generalization; it cannot recover unreported patient-level results.
Source Anchors
Shinkins and colleagues in BMJ distinguish intermediate, indeterminate, and uninterpretable results and explain why their handling must be reported. The BMJ diagnostic-tests series explains result-specific likelihood ratios. STARD 2015 provides reporting standards for diagnostic accuracy studies. None of these sources validate the invented counts above.
Keep reading
Don't stop at one method.
Good methods judgment comes from contrast. Read the neighboring guides, see where the assumptions diverge, and avoid treating every observational problem like it needs the same hammer.
The Expert-Reader Trap: Will Diagnostic Accuracy Travel to Routine Care?
A practical guide to checking whether the readers, training, and workflow in a diagnostic study match the setting where a test will be used.
Two Positive Tests Are Not Always Twice the Evidence
A worked example shows why repeating a correlated diagnostic test cannot automatically multiply the same likelihood ratio twice—and what joint pathway data reviewers should request.
Positive Predictive Value: Why a Good Test Can Still Produce Mostly False Alarms
A worked diagnostic-test example shows why positive predictive value changes with the tested population, even when sensitivity and specificity are held fixed—and why real-world accuracy may shift too.