🎯 Clinical Reasoning: Probabilistic Diagnosis

An educational tool for medical students and junior doctors to understand diagnostic reasoning, likelihood ratios, and clinical thresholds

The 3 Steps of Diagnosis

Every diagnostic decision follows the same pattern:

1 Estimate Pre-test Probability (Before testing)
"What's the chance this patient has [disease] before I examine them or order tests?"
Based on: prevalence in your setting, patient demographics, presenting complaint, risk factors
Example: "30-year-old with chest pain" → pre-test probability of MI is very low (~1%)
2 Apply Clinical Findings or Tests
Use findings from history (symptoms, timeline), examination (vital signs, physical findings), or investigations (bloods, imaging, bedside tests).
Each finding has a likelihood ratio (LR) that tells you how much to update your probability
Example: D-dimer negative (LR− 0.1) or ECG changes present (LR+ 5.0)
3 Calculate Post-test Probability (After findings)
"What's the chance this patient has [disease] now that I know they have [finding]?"
Compare this to your clinical thresholds to decide: treat immediately, investigate further, or rule out and reassure
Example: If post-test probability >50% → treat; if 10-50% → investigate; if <10% → safety-net

Understanding Likelihood Ratios

Definition: A likelihood ratio (LR) tells you how much a clinical finding changes the probability of disease.
  • LR+ = when the finding is present
  • LR− = when the finding is absent

How to interpret LRs:

LR+ Value Effect on Probability Clinical Meaning
>10 Large increase Strong evidence for disease
5–10 Moderate increase Good evidence for disease
2–5 Small increase Weak evidence for disease
1–2 Minimal change Finding doesn't help much
LR− Value Effect on Probability Clinical Meaning
<0.1 Large decrease Strong evidence against disease
0.1–0.2 Moderate decrease Good evidence against disease
0.2–0.5 Small decrease Weak evidence against disease
0.5–1.0 Minimal change Absence doesn't help much

The Mathematics (Don't Worry, We'll Visualise It!)

1 Convert pre-test probability to odds
Pre-test odds = probability ÷ (1 − probability)
Example: 10% probability = 0.10 ÷ 0.90 = 0.111 odds
2 Multiply odds by likelihood ratio
Post-test odds = pre-test odds × LR
Example: 0.111 × 0.25 (LR− for no abnormal vitals) = 0.028
3 Convert post-test odds back to probability
Post-test probability = odds ÷ (1 + odds)
Example: 0.028 ÷ 1.028 = 0.027 = 2.7%

Understanding with Natural Frequencies

Imagine 100 patients in your GP surgery who have lower respiratory infections, based on NICE guideline data 5–12% will be diagnosed and managed as community-acquired pneumonia.

Scenario: Using "No abnormal vital signs" as a rule-out test

LR− = 0.25 means the finding is 4 times less common in people WITH pneumonia than those WITHOUT.

Starting with 100 patients (example using 10% pre-test probability):
  • 10 have the disease
  • 90 do not have the disease
Disease present (10 patients)
Disease absent (90 patients)
After applying a finding with LR− = 0.25 (e.g., "No abnormal vital signs" for pneumonia):

This is a conceptual illustration showing approximately how findings distribute across diseased vs well patients. The exact numbers depend on the test's sensitivity and specificity.

Key insight: A negative finding (LR− 0.25) reduces probability from 10% → ~2.7% using Bayes' theorem.

Testing and Treatment Thresholds

Clinical decisions aren't just about calculating probability—you need to know what probability should trigger action. This is where thresholds come in.

The Threshold Model of Clinical Decision-Making
Every clinical decision involves comparing post-test probability against action thresholds. There are typically three zones:
  • Below test threshold: Disease unlikely enough to rule out without testing
  • Between test and treatment thresholds: Test to gain more information
  • Above treatment threshold: Disease likely enough to treat without further testing

Visual Representation

0%
Test Threshold
Treatment Threshold
100%
Rule Out Zone
No testing or treatment needed
Testing Zone
Perform diagnostic test
Treatment Zone
Treat empirically

How Thresholds Are Determined

Thresholds depend on several factors:

1 Severity of Disease
Life-threatening conditions (e.g., meningitis, MI, PE) have lower treatment thresholds—you're willing to treat at lower probability to avoid missing cases.
Example: Treat suspected meningitis at 10-15% probability; treat uncomplicated UTI at 50%+ probability
2 Risks of Treatment
Low-risk treatments (e.g., simple analgesia) can be given at lower probability thresholds. High-risk treatments (e.g., thrombolysis, immunosuppression) require higher certainty.
Example: Paracetamol for headache (minimal risk) vs anticoagulation for PE (bleeding risk)
3 Costs and Invasiveness of Testing
Low-cost, non-invasive tests (e.g., urine dipstick, ECG) have lower test thresholds. Expensive/invasive tests (e.g., angiography, biopsy) require higher pre-test probability to justify.
Example: ECG at 5% probability of MI vs cardiac catheterisation at 30%+ probability

Clinical Examples of Thresholds

Condition Test Threshold Treatment Threshold Rationale
Pulmonary Embolism ~2-5% ~15-20% Life-threatening; thresholds vary by Wells score and D-dimer availability
Strep Throat ~10-15% ~50-60% Usually self-limiting; antibiotics prevent rare complications
Pneumonia (primary care) ~5-10% ~30-40% Can be serious; antibiotics generally safe in short course
Bacterial Meningitis ~1-2% ~5-10% Rapidly fatal if untreated; antibiotics relatively safe

Interactive Likelihood Ratio Calculator

Practice calculating post-test probability using clinical findings from any condition.

Estimate based on:
• Prevalence in your setting
• Patient risk factors
• Clinical presentation

Practice Clinical Cases

Work through these realistic scenarios applying probabilistic reasoning and thresholds.

Case 1: Pneumonia - Low-risk presentation

Patient: 28-year-old non-smoker, 3-day history of cough with green sputum. No breathlessness. No comorbidities.

Examination: Temperature 37.2°C, HR 75, RR 16, SpO₂ 98%. Chest clear, no focal signs.

Your task:
  1. Estimate pre-test probability (hint: young, well, no focal signs = low)
  2. Apply LR− for "no abnormal vital signs" (0.25)
  3. Calculate post-test probability
  4. Decide: CXR? Antibiotics? Safety-net?
Show worked answer

Step 1: Pre-test probability ≈ 5% (young, well, no focal signs)

Step 2: Pre-test odds = 0.05 ÷ 0.95 = 0.053

Step 3: Post-test odds = 0.053 × 0.25 = 0.013

Step 4: Post-test probability = 0.013 ÷ 1.013 = 1.3%

Clinical decision: Pneumonia very unlikely. Safety-net advice, no CXR or antibiotics needed. Advise to return if symptoms worsen or don't improve in 3-5 days.

Case 2: Pneumonia - Moderate-risk presentation

Patient: 67-year-old with COPD, 5-day history of productive cough and breathlessness. Feels unwell.

Examination: Temperature 38.3°C, HR 98, RR 24, SpO₂ 94% on air. Crackles right base, dull to percussion.

Your task:
  1. Estimate pre-test probability (hint: elderly, comorbid, focal signs = moderate-high)
  2. Apply LR+ for temperature ≥38°C (3.21)
  3. Calculate post-test probability
  4. Decide: management plan?
Show worked answer

Step 1: Pre-test probability ≈ 25% (elderly, COPD, focal signs, systemically unwell)

Step 2: Pre-test odds = 0.25 ÷ 0.75 = 0.333

Step 3: Post-test odds = 0.333 × 3.21 = 1.069

Step 4: Post-test probability = 1.069 ÷ 2.069 = 51.7%

Clinical decision: Pneumonia likely. Consider point-of-care CRP if available. Start empirical antibiotics per local guidelines (example: amoxicillin 500mg TDS for 5 days—actual dosing varies by local antimicrobial guidance, allergy, severity). Arrange CXR if diagnostic uncertainty or high-risk features. Consider CRB-65 for admission decision. Safety-net with clear criteria to re-attend.

Case 3: Pneumonia - Using investigations to refine probability

Patient: 45-year-old smoker, 4-day cough, mild fever last night, feeling tired.

Examination: Temperature 37.8°C, HR 88, RR 18. Few scattered crackles bilaterally.

Point-of-care CRP: 15 mg/L

Your task:
  1. Estimate pre-test probability
  2. Apply LR− for CRP ≤20 mg/L (0.52)
  3. What's the post-test probability?
  4. How does this change management?
Show worked answer

Step 1: Pre-test probability ≈ 15% (smoker, mild systemic upset, equivocal chest signs)

Step 2: Pre-test odds = 0.15 ÷ 0.85 = 0.176

Step 3: Post-test odds = 0.176 × 0.52 = 0.092

Step 4: Post-test probability = 0.092 ÷ 1.092 = 8.4%

Clinical decision: CRP <20 mg/L substantially reduces pneumonia probability from 15% to 8%. This falls below typical treatment threshold. Reasonable to withhold antibiotics and provide safety-net advice. Advise return if fever persists >48h or symptoms worsen.

Case 4: Pulmonary Embolism - High-risk presentation

Patient: 52-year-old woman, 3-day history of right calf pain and swelling. Today developed sudden-onset breathlessness and pleuritic chest pain.

Examination: HR 105, RR 22, SpO₂ 94% on air. Right calf swollen and tender. Chest clear.

Wells Score: Clinical DVT signs (3 points) + tachycardia (1.5 points) + PE most likely diagnosis (3 points) = 7.5 points → "PE likely" (high risk)

Your task:
  1. Estimate pre-test probability: Wells >4 ("PE likely") corresponds to ~30-40% in ED populations
  2. Consider: Is D-dimer indicated at this level of risk?
  3. Decide: What is the appropriate next step?
  4. Reflect: What would happen in real-world practice?
Show worked answer

Pre-test probability analysis: Wells score 7.5 places this patient in the "PE likely" group. Studies show this corresponds to approximately 30-40% probability of PE in ED populations (some studies report 28-40% depending on exact cohort).

Is D-dimer indicated? No—and here's why:

  • Negative D-dimer cannot safely rule out PE at this baseline risk. Even with LR− 0.1: 35% → odds 0.54 × 0.1 → post-test ~5-6%, which remains above most rule-out thresholds for a life-threatening diagnosis
  • Positive D-dimer doesn't change management—you still need CTPA
  • Result: D-dimer adds delay without decision value

Recommended management (by-the-book): At 30-40% pre-test probability, proceed directly to CTPA, bypassing D-dimer. Consider early anticoagulation if clinically unstable or imaging delayed. This is standard practice for "PE likely" presentations.

⚠️ Real-world practice paradox:

Despite the above reasoning, D-dimer is still very often done in high-risk PE presentations. Why?

  • Reflexive ordering: Bloods (including D-dimer) are often sent at triage before formal risk stratification is complete
  • Workflow reality: By the time Wells score is calculated and the probability crystallizes, the D-dimer result already exists
  • Defensive documentation: Having "PE considered, D-dimer sent" in the notes is perceived as protective, even if it doesn't alter the decision
  • Cognitive disconnect: Clinicians run parallel tracks—formal reasoning ("needs CTPA") alongside system workflow ("send usual bloods")

Teaching point: In high-risk presentations, the D-dimer result is frequently ignored because it does not alter the decision to proceed to CTPA. This validates what juniors observe: the test gets done, but the decision was already made on clinical grounds.

Key learning: This case demonstrates why Bayesian reasoning sometimes leads you away from testing rather than toward it. At high pre-test probability, additional tests that won't cross thresholds should not be ordered—even though real-world practice may not always follow this principle. Understanding both the ideal reasoning and the practical reality is what makes you a better clinician.

Case 5: Acute MI - When strong findings exceed treatment threshold

Patient: 68-year-old with diabetes and hypertension. 45 minutes of severe central chest pain radiating to both arms and jaw. Sweating, nausea.

Examination: Pale, clammy. BP 95/60, HR 110. HS normal, chest clear.

ECG: ST elevation 3mm in leads II, III, aVF

Your task:
  1. Estimate pre-test probability (elderly, risk factors, classic history → high)
  2. Apply LR+ for ST elevation (13.1)
  3. Calculate post-test probability
  4. Consider: Do you need troponin before treatment, or is probability high enough to activate cath lab immediately?
Show worked answer

Step 1: Pre-test probability ≈ 40% (elderly, diabetes, hypertension, classic radiation pattern)

Step 2: Pre-test odds = 0.40 ÷ 0.60 = 0.667

Step 3: Post-test odds = 0.667 × 13.1 = 8.74

Step 4: Post-test probability = 8.74 ÷ 9.74 = 89.7%

Threshold analysis: Post-test probability (90%) is well above treatment threshold for STEMI (~80%). ECG finding alone provides sufficient certainty.

Clinical decision: This is STEMI. Immediate dual antiplatelet therapy, activate cath lab for primary PCI. Don't wait for troponin—probability already exceeds treatment threshold, and time-to-reperfusion is critical. This illustrates when further testing delays necessary treatment.

Case 6: Pulmonary Embolism – Low-risk presentation

Patient: 24-year-old female, sudden onset pleuritic chest pain. No calf swelling, no haemoptysis, no previous clots.

Examination: HR 72, SpO₂ 99% on air. Chest clear. Wells Score ≤4 ("PE Unlikely").

Investigation: D-dimer Negative.

Your task:
  1. Estimate pre-test probability (Wells "PE Unlikely" ≈ 10% prevalence)
  2. Apply LR− for Negative D-dimer (0.1)
  3. Calculate post-test probability
  4. Decide: CTPA or Discharge?
Show worked answer

Step 1: Pre-test probability ≈ 10%.

Step 2: Pre-test odds = 0.10 ÷ 0.90 = 0.111.

Step 3: Post-test odds = 0.111 × 0.1 = 0.011.

Step 4: Post-test probability = 0.011 ÷ 1.011 ≈ 1.1%.

Clinical decision: 1.1% is well below the test threshold for CTPA. You have successfully "ruled out" PE. Discharge with safety-netting.
Contrast this with Case 4: The same test (D-dimer) works here because the pre-test probability was low enough.

Case 7: Strep Throat – The 'Prevalence Trap'

Patient: 34-year-old father with a sore throat for 2 days. No cough.

Examination: Temp 37.1°C. Tonsils swollen with exudate (pus).

Context: Most adult sore throats are viral. Bacterial prevalence is low (~5-10%).

Your task:
  1. Estimate pre-test probability (~10%)
  2. Apply LR+ for Tonsillar Exudate (3.5)
  3. Calculate post-test probability
  4. Decide: Do you treat with antibiotics based on this finding?
Show worked answer

Step 1: Pre-test probability ≈ 10%.

Step 2: Pre-test odds = 0.10 ÷ 0.90 = 0.111.

Step 3: Post-test odds = 0.111 × 3.5 = 0.388.

Step 4: Post-test probability = 0.388 ÷ 1.388 ≈ 28%.

Clinical decision: Despite the "positive" finding of pus, the probability is only ~28%. This is often below the treatment threshold for antibiotics (usually >40-50% or Centor score 3-4).

The Lesson: This is the "Prevalence Trap" discussed in the Deep Dive tab. In low-prevalence settings, even good signs (LR+ 3.5) often fail to raise probability high enough to treat.

Making Sense of Diagnostic Tests

Why is this so confusing?
Medical students often struggle because there are two ways to look at test data:

  1. "Lab-facing" (Sensitivity/Specificity): How the test performs in a controlled group of known sick/well people. This is how tests are built.
  2. "Bedside-facing" (PPV/NPV): What the result means for the patient sitting in front of you. This is how tests are used.
The Golden Rule: Sensitivity and Specificity are fixed properties of the test. PPV and NPV change depending on the patient (prevalence).

The 4 Key Terms Explained Intuitively

Term The Question it Answers Clinical Role
Sensitivity "If the patient definitely has the disease, how often will the test be positive?" The Safety Net.
High sensitivity means few false negatives. Good for ruling out disease (SnNout).
Specificity "If the patient is definitely healthy, how often will the test be negative?" The Bouncer.
High specificity means few false positives. Good for ruling in disease (SpPin).
PPV
(Positive Predictive Value)
"My patient tested positive. What is the chance they actually have the disease?" Trust Level.
Changes based on prevalence. Often much lower than you think!
NPV
(Negative Predictive Value)
"My patient tested negative. What is the chance they are actually healthy?" Reassurance.
Usually very high in primary care because disease is rare.

⚠️ The "Prevalence Trap"

This is the most common mistake in diagnostic reasoning: Confusing Sensitivity with PPV.

Imagine a test for a cancer (1% prevalence in at risk group) that has 90% Sensitivity and 90% Specificity. Sounds like a great test, right?

If you get a Positive Result, do you have a 90% chance of cancer?

No! The chance is only roughly 8%.

Why? Let's look at 1,000 people:

10 Sick People
(1% Prevalence)

Test (90% Sens) catches:
9 True Positives
1 False Negative
990 Healthy People
(99% Healthy)

Test (90% Spec) correctly clears 891.
But 10% get it wrong:
99 False Positives

The Result: You have 9 true positives and 99 false positives.
PPV = 9 ÷ (9 + 99) ≈ 8.3%

Clinical Takeaway: Never interpret a test result without considering the Pre-test Probability (Prevalence). A positive test for a rare disease is likely a false positive unless the patient has specific symptoms that raise the pre-test probability.

Bridging the Gap: Why We Use Likelihood Ratios

If PPV changes every time the prevalence changes, how can we remember the numbers? We can't.

That is why Likelihood Ratios (LRs) are superior for clinical reasoning. LRs are calculated from Sensitivity and Specificity, so they (mostly) stay stable regardless of prevalence.

The "Bridge" Formulas:
LR+ = Sensitivity ÷ (1 − Specificity)
LR− = (1 − Sensitivity) ÷ Specificity

🛠️ "Lab to Bedside" Converter

Found a paper reporting Sensitivity/Specificity? Convert it to LRs here to see how useful the test actually is.

⚠️ Summary for the Wards:
  • Sensitivity/Specificity describe the test itself.
  • PPV/NPV describe the patient in front of you (and depend heavily on prevalence).
  • Likelihood Ratios are the bridge—they let you take your clinical gestalt (prevalence) and update it mathematically.
  • Tests don't make diagnoses—clinicians do.

Further Learning Resources

⚠️ Educational Tool Disclaimer: This is an educational resource to teach probabilistic reasoning. It is NOT a clinical decision tool or medical device. Always use clinical judgment, local guidelines (e.g., NICE NG63, BTS CAP guidelines), and safety-netting. When in doubt, seek senior advice or arrange face-to-face review.