Probabilistic Differential Diagnosis in Physiotherapy: A Practical Model

This article is an English summary of a teaching lecture by Philippe Tadger, originally delivered in Spanish (see the embedded video at the end). It walks through a probabilistic approach to differential diagnosis in musculoskeletal physiotherapy — how to turn a patient’s story into a ranked list of hypotheses, and how to update that ranking as new test results come in.

Why think in probabilities?

Clinical reasoning in physiotherapy is rarely black and white. A patient walks in with a story, a few symptoms, and a handful of signs — and the clinician must weigh several possible conditions at once. A probabilistic model makes that weighing explicit: each hypothesis has a pre-test probability, each test result nudges that probability up or down, and a decision to treat, refer, or reassure falls out of the updated number.

Patient script vs disease script — the semipermeable brick wall

The lecture starts from Maitland’s semipermeable brick wall model, which separates clinical reasoning into two compartments:

• The clinician side — anatomy, physiology, biomechanics, knowledge of dysfunction and pathology, metacognition (thinking about one’s own thinking), clinical skills, and the current scientific evidence. This is where diagnostic hypotheses are born.
• The patient side — the anamnesis, subjective account, beliefs and values, symptoms and signs, time course, comorbidities, behaviour, provocative and relieving factors, previous treatments and preferences. This is the raw material the clinician tries to explain.

Diagnosis becomes a matching exercise: how well does the patient script (everything you know about the person in front of you) overlap with the disease script (the typical presentation of a given condition)?

The body map: turning a story into data

A good body map captures each symptom separately — location, quality (sharp, cramping, fatigue-like), intensity (0–10), frequency (0–10, with 10 = constant), provocative factors, and relieving factors. Rather than one undifferentiated “back pain”, you end up with Symptom 1 = mechanical lumbar ache, 6/10, worst with prolonged sitting and Symptom 2 = subcostal visceral discomfort after fatty meals. The two patterns may not belong to the same condition at all — and that is exactly the kind of distinction a body map makes visible.

Pre-test probability: four tiers of likelihood

Once patient and disease scripts are side by side, each hypothesis is slotted into one of four tiers based on how well the scripts overlap:

• Very likely (> 67%) — the disease explains every finding, and every feature of the disease is present in the patient. Near-perfect overlap.
• Likely (34–66%) — the disease explains some findings, no strong rejection features, partial overlap.
• Unlikely (< 33%) — few shared features plus one or more rejection findings that argue against the condition.
• Red flag — low baseline probability, but the consequences of missing it are too serious to accept. Fractures, occult cancer, vascular pathology, and concussion fall here: you actively look for them with sensitive rule-out tests even when they feel improbable.

Updating probability with Fagan’s nomogram

Tests rarely give a yes/no answer. What they do give — when a decent study backs them — is a likelihood ratio (LR+ for a positive result, LR– for a negative result). Fagan’s nomogram is a simple graphical tool that takes a pre-test probability and an LR and spits out a post-test probability. A few worked examples from the lecture:

• Pre-test 40% + an LR near 1 → post-test still ~46%. The test adds nothing and can be dropped.
• Pre-test 20% + LR+ of 10 → post-test ~70%. A ‘probable’ becomes a ‘very probable’ — you can start treatment.
• Pre-test 40% + a good LR– → post-test ~7%. The condition is effectively ruled out; move on to other hypotheses.

The take-home is that a test with a strong LR+ earns its place for confirming a suspected diagnosis, and a test with a strong LR– earns its place for ruling out a dangerous one. Tests with LRs near 1 — on which so much time and money is spent — change almost nothing and should be questioned.

A worked cervical case

The lecture runs through a 25-year-old with chronic post-whiplash cervical pain, headaches and upper-limb dysaesthesias. Four hypotheses are ranked: mechanical cervical dysfunction, cervical radiculopathy, cervical myelopathy, and — as the red flag — odontoid fracture with alar ligament disruption. For each, the clinician pairs the most useful tests with their published LRs: Spurling and manual traction for radiculopathy (high LR+, confirmatory), upper-limb tension tests for ruling out, and the Sharp-Purser test (LR+ ≈ 17) for the alar ligament. One subtle point: the same test cannot legitimately carry the same LR for two different conditions — if your table shows that, it is a sign the evidence is being copy-pasted without scrutiny.

Takeaways

• Write out the patient script and compare it to several disease scripts before reaching for tests.
• Rank your hypotheses into four tiers, and keep the red-flag category even when the probability is low.
• Pick tests that actually move the probability — strong LR+ to confirm, strong LR– to rule out.
• Use a simple tool (Fagan’s nomogram) to make the update explicit rather than implicit.
• Be sceptical of evidence tables where the same test carries identical LRs across conditions.

Watch the original lecture (Spanish)

Original title: “Como hacer Diagnóstico Diferencial en Fisioterapia usando un Modelo Probabilístico” — Philippe Tadger, on YouTube.