Evidence-Based Medicine · Course for Techies

Techie Summary: EBM in the Language You Already Speak

A fast, jargon-translated map of all nine lessons — for readers who think in systems, tests, and data pipelines rather than clinics.

This course teaches evidence-based medicine (EBM) — how clinicians judge whether a piece of medical research is trustworthy and relevant. If you have no clinical background but do have a technical one, the fastest on-ramp is mapping each concept to something you already know. That's what this page does. It isn't a replacement for the nine lessons — it's the cheat sheet you skim first, and return to whenever a lesson's jargon gets thick.

One-line version of the whole course: a clinical question is a search query, a research paper is a system whose reliability you have to test before you trust its output, and statistics are just the error bars and significance tests you'd want on any experiment before shipping a decision based on it.

The Medicine → Tech Dictionary

Medical term	Techie translation
PICO question	A well-formed search query / schema: Population, Intervention, Comparison, Outcome — the four required fields before you query the literature
Randomized controlled trial (RCT)	An A/B test with random assignment to control for confounders
Cohort / case-control study	Observational log analysis — you didn't control assignment, so you have to work harder to rule out confounders
Confounder	A hidden third variable correlated with both your "feature" and your "outcome," producing a spurious correlation
Sensitivity / specificity	Recall / (1 − false positive rate) for a diagnostic test, evaluated like a binary classifier
Likelihood ratio	A Bayesian update multiplier — same math as a spam filter updating "P(spam)" after seeing one more signal
ARR / RRR / NNT	Absolute vs. relative improvement, and "cost per unit of benefit" — the antidote to a misleading percentage in a headline
Confidence interval	Error bars on the effect estimate; a CI crossing "no effect" is like an experiment result that isn't statistically distinguishable from noise
P-value	Output of a null-hypothesis significance test — not "probability the treatment works," just "how surprising this data would be if it didn't"
Meta-analysis / forest plot	Combining several underpowered experiments into one weighted-average result with a recalculated confidence interval
Heterogeneity (I²)	Variance across experiments beyond what sampling noise explains — a signal that you may be averaging apples and oranges
Publication bias / funnel plot	Survivorship bias in your dataset — "boring" negative results are underrepresented because they're less likely to get published
GRADE (certainty vs. recommendation)	Test-coverage score (how much you trust the data) vs. the actual go/no-go ship decision (which also weighs cost and risk tolerance)
Hazard ratio / Kaplan-Meier curve	Time-to-event analysis — like a churn curve, comparing how fast two cohorts hit an "event" over time, not just whether they eventually do

The Course, Lesson by Lesson (Fast Version)

Lesson 1 — Start with the patient (the "user"), not the paper. History and exam are basically requirements-gathering before you touch any literature.
Lesson 2 — Turn a vague worry into a PICO query, then run it against PubMed like a structured database search.
Lesson 3 — Code-review an RCT: check validity (was it built right?) before results (what did it output?) before applicability (does it apply to your specific case?).
Lesson 4 — Meta-analyses pool multiple trials into one more-confident estimate; forest plots and funnel plots are how you sanity-check that pooling.
Lesson 5 — Diagnostic test performance is classifier evaluation: sensitivity/specificity, PPV/NPV, likelihood ratios, and ROC curves for choosing a cutoff.
Lesson 6 — Prognosis and harm studies are usually observational (no random assignment possible), so confounding control becomes the central engineering problem.
Lesson 7 — Clinical guidelines are a "recommendation engine" built on graded evidence (GRADE) — separating data quality from the final decision.
Lesson 8 — A skeptic's deep-dive into what p-values, significance, and "limitations" sections actually claim — and where they get misused (p-hacking, surrogate endpoints).
Lesson 9 — One full patient case run through the entire pipeline end to end, tying back to Lesson 1.

Why this matters if you're not a clinician: the whole point of EBM is the same reason you write tests before deploying code — a plausible-sounding claim ("this drug works," "this feature increases retention") still needs its methodology checked before you trust the number attached to it.

Recommended Path

If you're short on time: read Lesson 1 (five minutes), then Lesson 3 (the core appraisal skill), then come back to this page whenever a term in Lessons 4–8 needs translating. Lesson 9 is worth doing last, in full, since it's the only lesson that shows the whole pipeline running on one real case.