01Principles and purposeThe professional or clinical skill and the decisions it supports.
Diagnostic reasoning is an iterative cycle rather than a one-way march from symptom to label. Information gathering creates a problem representation; that representation activates plausible mechanisms; targeted evidence shifts their relative probability; and the working diagnosis guides action and further observation. The cycle remains open because diseases evolve, tests are imperfect and several conditions may coexist. A useful conclusion therefore includes probability, consequence and the next decision rather than merely a disease name.
Bayes' theorem formalises updating. Pre-test odds equal probability divided by one minus probability. Multiply those odds by the likelihood ratio for the observed result to obtain post-test odds, then divide odds by one plus odds to return to probability. Likelihood ratios are built from sensitivity and specificity for dichotomous tests: LR+ equals sensitivity divided by one minus specificity, while LR− equals one minus sensitivity divided by specificity. The calculation is only as sound as the chosen pre-test probability and the population in which accuracy was estimated.
Test selection should anticipate an action boundary. A highly sensitive test may be useful after a negative result when the population and threshold match, but 'sensitive means rule out' is not a universal incantation. Specificity, indeterminate zones, continuous values, timing, specimen quality and downstream confirmation all matter. A result near a cut-off is not biologically discontinuous. Serial tests are correlated when they measure the same process, so multiplying their likelihood ratios as if independent can overstate certainty.
Errors arise from cognition and systems. Anchoring fixes attention on an early frame; premature closure stops the search; confirmation bias privileges supportive evidence; base-rate neglect ignores prior probability. Yet simply naming biases does not reliably repair them. Operational checks are stronger: ask what else fits, seek disconfirming evidence, compare the raw result with the report, review whether the patient matches the validation population, specify ownership of pending results and reopen the diagnosis when the clinical course violates predictions.
Key points
- Define the decision before ordering a test: rule out a dangerous possibility, confirm a treatable diagnosis, stage disease, select treatment or monitor change.
- Estimate pre-test probability from the patient's pattern and relevant population; predictive values cannot be transported unchanged when prevalence differs.
- Convert probability to odds, multiply by the likelihood ratio, then convert post-test odds back to probability when a quantitative update is needed.
- Sensitivity and specificity describe performance conditional on disease status; positive and negative predictive values describe what a result means in the tested population.
- A reference standard can be imperfect, unavailable or applied selectively, creating misclassification, verification and spectrum bias in accuracy estimates.
- Test thresholds link probability to action: below a testing threshold observe or stop, between thresholds gather evidence, and above a treatment threshold act when benefits justify it.
- Use a diagnostic pause for discordance, deterioration or failed treatment: restate the problem, inspect raw data, reconsider alternatives and assign follow-up.
02Situations and prioritiesThe context, relevant information and actions that matter most.
The probability before the new test combines setting, prevalence and patient-specific evidence; a referral clinic estimate may not fit community screening.
A likelihood ratio states how much more or less common the observed result is with disease than without it, moving odds rather than adding probability points.
The same post-test probability can justify treatment, further testing or observation depending on disease harm, treatment burden and patient preference.
Accuracy changes when severity, comorbidity, competing disease or operator skill differs from the study, even if the assay name is identical.
Two tests responding to the same biological signal do not provide independent evidence; naive sequential multiplication can create false confidence.
A result that does not explain physiology or trajectory requires verification and reframing, including the possibility of multiple simultaneous diagnoses.
03Assessment and interpretationHow to gather information, assess the situation and recognise uncertainty.
Consider the information, its meaning and its limitations before deciding what follows.
- 01
Pre-test probability estimate - Why
- Anchor interpretation to the patient and setting before the result is known.
- Interpretation and limitations
- Use a validated clinical model when applicable or state a reasoned range; retrospective estimates after seeing the result are vulnerable to hindsight.
- 02
Sensitivity and specificity appraisal - Why
- Assess how the index test separated reference-positive and reference-negative groups.
- Interpretation and limitations
- Check threshold, confidence intervals, patient spectrum, blinding and reference standard; these measures do not directly provide an individual patient's probability.
- 03
Likelihood-ratio update - Why
- Translate a result into a change from pre-test to post-test odds.
- Interpretation and limitations
- Use the LR for the actual result interval where possible, not merely positive or negative; avoid independent multiplication of correlated results.
- 04
Predictive-value appraisal - Why
- Describe the proportion of test-positive or test-negative people with the corresponding reference state.
- Interpretation and limitations
- PPV and NPV depend on prevalence and study selection, so transport to a different clinical setting requires recalibration.
- 05
Reference-standard review - Why
- Judge whether disease status was classified independently and accurately.
- Interpretation and limitations
- An imperfect reference, differential verification or failure to verify test-negative participants can bias apparent sensitivity and specificity.
- 06
Clinical utility check - Why
- Determine whether testing changes decisions or patient outcomes rather than only classification.
- Interpretation and limitations
- Consider harms from sampling, false results, overdiagnosis, delay and downstream procedures alongside accuracy and turnaround time.
04Worked approachesCases with ordered reasoning, an action and a check of the outcome.
01Worked exampleUpdate probability with a negative testInputs: pre-test probability 30%; a negative result from a test with sensitivity 90% and specificity 80%; decide whether the result crosses a 5% action threshold.+
- 1Convert pre-test probability to odds: 0.30 divided by 0.70 equals 0.4286.
- 2Calculate LR−: one minus sensitivity, 0.10, divided by specificity, 0.80, equals 0.125.
- 3Multiply: post-test odds equal 0.4286 times 0.125, which is 0.0536.
- 4Convert back: 0.0536 divided by 1.0536 equals 0.0509, or about 5.1%; the result sits just above rather than below the stated 5% threshold.
- 5Conclude that the negative test narrowly fails the predefined threshold, so further evidence or observation is needed rather than calling disease excluded.
- 6Verify with 1,000 people at 30% prevalence: 300 have disease, giving 270 true positives and 30 false negatives at 90% sensitivity; 700 do not have disease, giving 560 true negatives and 140 false positives at 80% specificity. Of the 590 negative results, 30 have disease: 30/590 = 5.08%, agreeing with the odds calculation and remaining just above the hypothetical 5% action threshold.
02Test-selection pathwayAsk what the result can changeA stable patient has an intermediate probability of a condition and several possible tests.+
- 1Define the immediate decision and the probability threshold that would change it.
- 2Compare candidate tests for applicable accuracy, harms, delay, indeterminate results and downstream consequences.
- 3Choose the least burdensome test likely to move probability across a decision threshold, or do not test if none can.
- 4Pre-plan how positive, negative and indeterminate results will be acted upon and who will review them.
03Diagnostic timeoutReopen a failing explanationThe patient deteriorates despite treatment aimed at the working diagnosis.+
- 1Restate the problem representation using current physiology and timeline rather than copying the admission label.
- 2Inspect original observations, raw test data, timing, sampling and treatment exposure for technical or interpretive error.
- 3Generate serious alternatives, complications and coexisting diagnoses, deliberately seeking evidence that contradicts the current frame.
- 4Escalate care by physiology, obtain targeted evidence, document uncertainty and set a short reassessment interval.
05Feedback, follow-up and evidenceReview outcomes, seek feedback and identify what to improve.
- Record the pre-test estimate or qualitative category before receiving the test result when it materially affects interpretation.
- Track every pending or indeterminate result to a named reviewer, patient communication plan and action deadline.
- Recalculate probability when prevalence, clinical state or test timing changes rather than carrying forward an obsolete estimate.
- Audit false-positive and false-negative outcomes in the local pathway, including who was never verified by the reference standard.
- Use follow-up as diagnostic evidence: define the predicted trajectory and reopen assessment when observation or treatment response violates it.
06Special situationsVariants, exceptions and circumstances that change the usual approach.
Odds are the multiplying scale
Likelihood ratios multiply odds, not probabilities. Converting correctly prevents impossible arithmetic and makes sequential evidence transparent.
A threshold is a decision construct
Biology does not jump at a laboratory cut-off. Costs, harms and preferences determine what to do near the boundary.
Verification can distort accuracy
If only positive index tests receive the definitive reference procedure, false negatives remain hidden and sensitivity may appear better than it is.
Treatment response is weak when nonspecific
Improvement after fluids, analgesia or broad therapy may fit several mechanisms; response supports a diagnosis only when predicted and reasonably specific.
No result is context-free
Time from onset, prior treatment, specimen quality, operator and assay threshold can change the meaning of the same printed value.
07Common pitfallsFrequent interpretation and management errors.
- 01
Ordering a test because it is available without defining the decision its result could change.
- 02
Interpreting positive predictive value as a fixed property that transfers unchanged from a specialist clinic to population screening.
- 03
Adding likelihood ratios directly to probability or multiplying correlated test results as if they were independent.
- 04
Calling a disease ruled out when post-test probability remains above the threshold for safe non-action.
- 05
Accepting the report conclusion without checking specimen, threshold, reference method, confidence interval or clinical fit.
- 06
Responding to deterioration by repeating the same tests while leaving the original problem representation untouched.