A Point-of-Care HbA1c Device: Predictive Values, Budget Reality and the Rule-In/Rule-Out Debate
A regional authority is considering a point-of-care HbA1c device for opportunistic diabetes screening in 12 primary care centres. In a validation study, 1,000 adults aged 40 and over gave a finger-prick sample for the device and, the same day, a venous sample for the reference laboratory (the reference standard). Local diabetes prevalence in this age group is 12%. Of the 120 people classified as diabetic by the reference method, 98 tested positive on the device; of the 880 classified as non-diabetic, 88 tested positive. The device costs $6 per test and a laboratory HbA1c costs $22. The authority has $150,000 for the first year and expects to screen 25,000 people. One clinician group wants the device used as a rule-out test; another wants it as a rule-in test before starting treatment.
1. Construct the 2×2 table and calculate sensitivity, specificity, positive and negative predictive values and overall accuracy. Explain to a non-clinical manager, in plain language, why the predictive values differ so much from the sensitivity.
Read the worked answer
2×2 table
| Reference: diabetic | Reference: not diabetic | Total | |
|---|---|---|---|
| Device positive | 98 | 88 | 186 |
| Device negative | 22 | 792 | 814 |
| Total | 120 | 880 | 1,000 |
- Sensitivity = 98/120 = 81.7%
- Specificity = 792/880 = 90.0%
- PPV = 98/186 = 52.7%
- NPV = 792/814 = 97.3%
- Accuracy = 890/1,000 = 89.0%
Plain-language explanation. Sensitivity describes the test under laboratory conditions: of every 100 people who truly have diabetes, the device finds about 82 and misses 18. That sounds reassuring. But the manager's real question is different: a patient has just tested positive - what is the chance the result is right? Here it is barely better than a coin toss, and the reason is arithmetic, not test quality. Diabetes is uncommon (12%), so the pool of healthy people is large. A 10% false-positive rate applied to 880 healthy people generates 88 false alarms, almost as many as the 98 genuine detections. When the diseased group is small and the healthy group is large, false positives outnumber true positives. Conversely, a negative result is trustworthy, because 97% of negatives really are non-diabetic. The lesson: predictive values belong to the setting, not to the device, and they shift as prevalence shifts.
2. Applied to 25,000 people at 12% prevalence, estimate true cases detected, cases missed, false positives and total first-year cost including confirmatory testing. Is the budget sufficient, and what would you advise?
Read the worked answer
Yield in 25,000 people
- True diabetics = 25,000 × 0.12 = 3,000; non-diabetics = 22,000
- Detected = 3,000 × 0.817 ≈ 2,450; missed ≈ 550
- False positives = 22,000 × 0.10 = 2,200
- Total device positives = 4,650 (false positives are 47% of them)
Costs
- Devices: 25,000 × $6 = $150,000
- Confirmatory laboratory HbA1c on every positive: 4,650 × $22 = $102,300
- Total = $252,300 against a $150,000 envelope - about 68% over budget.
Advice. The budget cannot deliver the plan as written, and confirmatory testing is not a discretionary extra: without it, roughly half of the people labelled diabetic would be mislabelled and possibly started on treatment. Options, in order of preference: (1) restrict testing to people at higher prior risk (age, BMI, family history, previous gestational diabetes), which raises prevalence and therefore PPV while shrinking the screened population; (2) negotiate a device price and a bulk laboratory rate; (3) phase implementation across centres with an explicit evaluation, rather than launching everywhere; (4) plan for the downstream costs of the newly diagnosed - extra consultations, medication, retinal screening - which are not in the envelope at all. Reducing the screened population to about 14,800 would fit the current budget, but the authority should choose that openly rather than discover it halfway through the year.
3. The two clinician groups disagree about rule-in versus rule-out use. How does prevalence change the interpretation, and what study and reporting limitations should temper the authority's decision?
Read the worked answer
Rule-out versus rule-in. Rule-out demands near-perfect sensitivity and a very high NPV, because a false negative sends a diseased person away untreated. Rule-in demands a high PPV, because a false positive commits someone to lifelong treatment, monitoring and cost. At 12% prevalence this device is a defensible rule-out adjunct - NPV 97.3%, so a negative result is reassuring, though 550 of 3,000 cases would still be missed, so it cannot be the only test in a symptomatic patient. It is not a rule-in test: a PPV of 52.7% means that treating every positive would treat nearly one person without disease for every one with it. Every positive must be confirmed.
How prevalence moves PPV. Recomputing with the same sensitivity and specificity: at 3% prevalence, of 1,000 people, 30 are diseased → 24.5 true positives, and 970 healthy → 97 false positives; PPV = 24.5/121.5 ≈ 20%. At 30% prevalence, 300 diseased → 245 true positives, 700 healthy → 70 false positives; PPV = 245/315 ≈ 78%. The same device is a mediocre test in the general population and a good test in a high-risk clinic.
Limitations to flag. Single-site validation in volunteers (spectrum bias), no confidence intervals reported, unclear whether all participants received the reference test (verification bias), unknown field performance with non-laboratory staff and variable finger-prick technique, no data on repeat testing or on the reproducibility of results near the cut-off. Request cluster-level field evaluation before a regional rollout.