Biostatistics & Population Health

Decision trees in clinical decision-making

Clinical Overview and When to Suspect Decision Tree Utility

— Decision nodes (squares): clinician-controlled choices (treat vs. observe vs. test)

— Chance nodes (circles): probabilistic events (disease present/absent, treatment success/failure)

— Terminal nodes (triangles): outcomes assigned utilities (QALYs, costs, mortality)

— Branches carry probabilities that must sum to 1.0 at each chance node

— Discrete, short-horizon problem (testing strategy, single treatment choice)

— Outcomes are time-independent or occur within a bounded period

— Probabilities and utilities can be estimated from literature

— Contrast with Markov models, used when patients transition repeatedly between health states over long horizons (e.g., chronic kidney disease progression)

— Cost-effectiveness analyses (screening colonoscopy vs. FIT)

— Threshold analysis (test-treat thresholds based on disease probability)

— Shared decision-making aids (anticoagulation in atrial fibrillation)

— Quality improvement and resource-allocation discussions

Decision trees are structured graphical models that map sequential clinical choices, chance events, and outcomes to support evidence-based decisions under uncertainty

Core anatomy:

When to suspect a decision-tree analysis is the right framework:

Clinical contexts on Step 3:

Expected value calculation is the engine: at each chance node, multiply probability × outcome value; fold back ("averaging out and folding back") from right to left to choose the highest-value branch at each decision node

Board pearl: Decision trees evaluate expected utility, not guaranteed outcomes — a "correct" decision can still yield a bad result in an individual patient; this distinguishes decision quality from outcome quality and is a recurring Step 3 ethics/QI theme

Key distinction: Decision tree = one-shot branching pathway; Markov model = recurring cyclic transitions; sensitivity analysis tests how robust the recommendation is to changes in input probabilities or utilities

Presentation Patterns and Key History

— Stem gives pretest probability and asks whether to test, treat empirically, or do neither

— Below testing threshold → no test, no treat

— Between thresholds → test, then act on result

— Above treatment threshold → treat without testing (e.g., classic angina in a 70-year-old smoker — treat, don't stress-test first)

— Two strategies compared with costs and QALYs

— Calculate ICER = (Cost_A − Cost_B) / (QALY_A − QALY_B)

— Threshold typically cited as $50,000–$150,000/QALY in the US

— Patient with borderline indication (e.g., CHA₂DS₂-VASc 1, prostate cancer screening at age 70)

— Tree clarifies tradeoffs between bleeding vs. stroke, overdiagnosis vs. mortality benefit

— Baseline disease probability (prevalence in this population)

— Test characteristics (sensitivity, specificity, LR+/LR−)

— Treatment efficacy and harm rates

— Patient values/preferences (a 75-year-old who prioritizes quality over longevity)

— Perfect health = 1.0, death = 0.0

— Stroke with disability ≈ 0.3–0.5; major bleed ≈ 0.6–0.7

— These weights drive the "right" branch

Decision trees "present" on Step 3 in several recognizable stem formats — pattern recognition saves time

Test-treatment threshold stems:

Cost-effectiveness stems:

Shared decision-making stems:

Key history elements the stem will provide:

Step 3 vignettes often embed utility weights:

Step 3 management: When a vignette lists multiple management options with explicit probabilities and outcomes, do not pick the option with the best possible outcome — pick the one with the highest expected value (probability-weighted average). This is the single most common decision-analysis trap on the exam

Board pearl: "Watchful waiting" is a legitimate branch and frequently the correct answer when test harms or treatment harms outweigh modest probability gains

Physical Exam Findings (and Probability Revision)

— LR+ = sensitivity / (1 − specificity)

— LR− = (1 − sensitivity) / specificity

— Posttest odds = pretest odds × LR

— LR >10 or <0.1 → large, often conclusive shift

— LR 2–5 or 0.2–0.5 → small to moderate shift

— LR ≈ 1 → useless test/finding

— S3 gallop for heart failure: LR+ ≈ 11

— Calf tenderness alone for DVT: LR+ ≈ 1.1 (minimally useful — explains why Wells score + D-dimer dominate the DVT decision tree)

— Murphy sign for cholecystitis: LR+ ≈ 2.8

— Start at decision node "test or not"

— Chance node splits into disease+/disease−

— Each splits into test+/test− using sensitivity/specificity

— Terminal utilities assigned by true/false positive/negative status

— Testing threshold = probability below which testing causes more harm than benefit

— Treatment threshold = probability above which empiric treatment beats testing-then-treating

— Both depend on test characteristics AND on the harm-to-benefit ratio of treatment

In decision-tree problems, the "physical exam" equivalent is Bayesian probability revision — how findings update pretest to posttest probability

Likelihood ratio framework:

Examples of high-LR exam findings that feed decision trees:

Building the branch:

Threshold logic in exam reasoning:

A test with poor specificity raises the testing threshold (more false positives → more downstream harm)

A treatment with high toxicity raises the treatment threshold (need more certainty before pulling the trigger)

Board pearl: The threshold approach (Pauker & Kassirer) is the conceptual backbone of nearly every Step 3 "should you order this test?" question — internalize that a positive test is only useful if it would change management

Key distinction: Sensitivity/specificity are test properties; PPV/NPV depend on prevalence and therefore differ across populations even with the identical test

Diagnostic Workup — Building and Populating the Tree

— Step 1: Frame the clinical question with mutually exclusive, collectively exhaustive options

— Step 2: Identify the time horizon (acute event vs. lifetime)

— Step 3: Draw decision nodes (squares) for clinician choices

— Step 4: Add chance nodes (circles) for probabilistic events; branches must sum to 1.0

— Step 5: Assign terminal utilities (QALYs, life-years, costs, or composite)

— Step 6: Fold back — calculate expected value at each chance node, choose max at each decision node

— Probabilities: systematic reviews, meta-analyses, registry data, local epidemiology

— Utilities: standard gamble, time trade-off, or validated catalogs (e.g., CDC HRQOL, EQ-5D)

— Costs: Medicare reimbursement schedules, Red Book pricing, micro-costing studies

— Pretest probability ≈ 80% with classic symptoms

— Empiric treatment branch: high cure rate, modest C. diff/resistance risk

— Test-first branch: 1–2 day delay, possible undertreatment

— Folding back typically favors empiric treatment — exactly why guidelines recommend it

Step-by-step construction (the way Step 3 expects you to reason):

Data sources for inputs:

Worked example — empiric antibiotics vs. culture-guided therapy for suspected UTI in a young woman:

CCS pearl: In CCS cases involving common, high-pretest-probability conditions (uncomplicated cystitis, strep pharyngitis in a child with Centor 4, classic shingles), ordering confirmatory tests before treating costs you simulated time and points — the decision-tree logic mirrors real CCS scoring

Board pearl: Always verify that probabilities at any single chance node sum to exactly 1.0; an exam answer choice with branches summing to 0.9 or 1.1 is wrong by construction

Diagnostic Workup — Sensitivity Analysis and Model Validation

— One-way: vary a single parameter across its plausible range, observe whether the preferred strategy changes

— Two-way: vary two parameters simultaneously, often visualized on a 2D plot with a "threshold line"

— Probabilistic (Monte Carlo): assign distributions to all parameters, run thousands of simulations, report % of iterations favoring each strategy

— Tornado diagram: ranks parameters by impact on the result — widest bars = most influential variables

— Example: if anticoagulation is preferred when annual stroke risk >1.7%, that 1.7% is the threshold and it should drive your CHA₂DS₂-VASc cutoff

— Face validity (does it match clinical intuition?)

— Internal validity (math correct, probabilities sum to 1?)

— External validity (predictions match observed real-world outcomes?)

— Cross-validation against published trials or registries

— Ignoring discounting of future costs and benefits (standard 3%/year in US analyses)

— Double-counting (e.g., counting a stroke both as event and as utility decrement without aligning time frames)

— Using disease-specific rather than all-cause mortality when comparing screening strategies — biases toward the screening arm

— Conflating efficacy (RCT) with effectiveness (real world)

A decision tree's recommendation is only as trustworthy as its inputs — sensitivity analysis stress-tests robustness

Types of sensitivity analysis:

Threshold value: the input value at which two strategies have equal expected utility — crossing it flips the recommendation

Validation steps:

Common pitfalls Step 3 tests:

Board pearl: If a sensitivity analysis shows the preferred strategy doesn't change across the plausible range of any input, the recommendation is robust — this is the gold-standard endorsement for guideline adoption

Key distinction: Deterministic sensitivity analysis gives a "threshold"; probabilistic analysis gives a "probability of cost-effectiveness" — the latter is now preferred by ISPOR and CHEERS reporting standards

Risk Stratification — Threshold Approach to Testing and Treatment

— Zone 1 (below testing threshold): disease so unlikely that testing harms exceed benefits → do not test, do not treat

— Zone 2 (between thresholds): test, then treat based on result

— Zone 3 (above treatment threshold): disease so likely that empiric treatment dominates → treat without testing

— Ptt = [(1 − Sp) × H] / [(1 − Sp) × H + Sn × B]

— where H = harm of treating disease-free patients, B = benefit of treating diseased patients

— Prx = [Sp × H] / [Sp × H + (1 − Sn) × B + ...] (full form includes test risks)

— Healthy 25-year-old with atypical chest pain → below testing threshold for stress test

— 55-year-old diabetic with exertional pressure → between thresholds → stress test

— 70-year-old with classic crescendo angina → above treatment threshold → cath/treat, skip stress test

— Safer treatment → lowers treatment threshold (easier to justify empiric Rx)

— More dangerous test → raises testing threshold

— Higher test sensitivity → lowers testing threshold (won't miss disease)

— Higher test specificity → lowers treatment threshold (positive test is trusted)

The Pauker-Kassirer threshold model is the single most testable decision-analysis framework on Step 3

Three zones along the pretest-probability axis (0% → 100%):

Calculating the testing threshold (Ptt):

Calculating the treatment threshold (Prx):

Clinical anchors:

What shifts the thresholds:

Step 3 management: When a vignette gives you a pretest probability and asks for next step, mentally place the patient in one of the three zones before choosing — this prevents the "always order one more test" reflex that costs points on both MCQ and CCS

Board pearl: Pulmonary embolism workup is the canonical example — PERC rule defines the below testing threshold zone in low-risk patients

Pharmacotherapy — Applying Decision Trees to Drug Selection

— Decision node: warfarin vs. DOAC vs. no anticoagulation vs. left atrial appendage occlusion

— Chance nodes: annual stroke risk (driven by CHA₂DS₂-VASc), major bleed risk (HAS-BLED), intracranial hemorrhage rate

— Utilities: stroke with disability ≈ 0.3, GI bleed ≈ 0.7, well on anticoagulant ≈ 0.99

— Folding back: DOACs typically dominate warfarin (similar efficacy, ~50% less ICH, no INR monitoring) — except in mechanical valves and moderate-severe mitral stenosis

— ASCVD 10-year risk <5% → below threshold, no statin

— 5–7.5% → shared decision-making zone (use risk enhancers)

— ≥7.5% → above threshold, statin recommended

— Pooled Cohort Equations feed the chance node

— Tree branches on cardiovascular vs. renal vs. cost priorities

— SGLT2 inhibitor preferred when HFrEF, CKD, or established ASCVD present (high-utility branch)

— GLP-1 RA preferred for ASCVD without HF and for weight benefit

— Metformin remains first-line absent these comorbidities

— NNT = 1 / absolute risk reduction

— NNT 20 with a benign drug is favorable; NNT 100 with a toxic drug rarely is

Decision trees rigorously evaluate competing pharmacotherapies by integrating efficacy, toxicity, cost, and patient utility

Classic Step 3 application — anticoagulation in nonvalvular atrial fibrillation:

Statins for primary prevention:

Glycemic control in T2DM:

Number needed to treat (NNT) is the bedside instantiation of decision-tree folding:

Step 3 management: When two drugs have similar efficacy, the decision tree pivots on adverse events × patient utility weights — a patient who fears bleeding above all else should not be given the drug with double the GI-bleed rate, regardless of population-level expected value

Board pearl: Cost-effectiveness alone never overrides patient preference in a shared-decision framework — the tree informs, the patient decides

Procedures and Invasive Management — Expanded Decision Analysis

— Decision node: surgery vs. medical management vs. minimally invasive alternative

— Chance nodes: perioperative mortality, complication rates, long-term benefit, recurrence

— Terminal utilities must include perioperative QALY decrement (typically 1–3 months of reduced health utility)

— Benefit only if perioperative stroke/death rate <3% at the operating center

— Above that, medical therapy (statin, antiplatelet, BP control) dominates

— Threshold is center- and surgeon-specific, embedding quality data into the tree

— SYNTAX score serves as the pretest probability input

— High SYNTAX (>33) + diabetes → CABG branch dominates (FREEDOM trial)

— Low SYNTAX → PCI competitive

— Colonoscopy: highest sensitivity, highest procedure risk (perforation 1/1000), 10-year interval

— FIT: lower sensitivity per round but annual repetition compensates

— Cost-effectiveness analyses generally show all three within acceptable ICER bands

— Clinical Frailty Scale ≥5 dramatically raises perioperative chance-node probabilities

— Often flips an otherwise-favorable surgery branch into the medical-management branch

— Step 3 increasingly tests functional status over chronologic age

Procedural decisions are the highest-stakes application of decision trees because outcomes are largely irreversible

Framework for "should we operate?" trees:

Carotid endarterectomy for asymptomatic stenosis — classic exam vignette:

CABG vs. PCI in multivessel disease:

Screening colonoscopy vs. FIT vs. CT colonography:

Frailty as a tree modifier:

CCS pearl: Before ordering a procedure in a CCS case, mentally run the tree — if the patient is unlikely to benefit (advanced dementia, metastatic cancer with poor prognosis, severe frailty), the right move is goals-of-care conversation as the next order, not the procedure itself

Board pearl: Informed consent for a procedure should include the probabilities and outcomes the decision tree quantified — not just qualitative risk language

Special Populations — Elderly and Renal/Hepatic Impairment

— Shorter remaining life expectancy compresses the time horizon — long-latency benefits (e.g., cancer screening) lose value

— Lead time to benefit must be shorter than life expectancy for screening or preventive therapy to make sense

— Examples: colon cancer screening benefit lead time ≈ 7–10 years; mammography ≈ 10 years; statins for primary prevention ≈ 2–5 years

— A 78-year-old with 6-year life expectancy gains little from new screening colonoscopy but may still benefit from statin therapy

— Death from other causes becomes a major terminal node

— Disease-specific interventions yield diminishing marginal benefit

— Aggregate comorbidity indices (Charlson, ePrognosis) recalibrate baseline probabilities

— DOAC dosing thresholds (apixaban 2.5 mg BID if 2 of: age ≥80, weight ≤60 kg, Cr ≥1.5)

— Avoid dabigatran if CrCl <30; avoid rivaroxaban/edoxaban if CrCl <15

— Contrast nephropathy risk reshapes the imaging branch (favor MRI or non-contrast CT)

— Avoid acetaminophen >2 g/day in cirrhosis

— Statin selection: pravastatin and rosuvastatin preferred (less hepatic metabolism)

— DOACs contraindicated in Child-Pugh C

— Each added medication ~10% increase in adverse event probability

— Beers and STOPP/START criteria operationalize this in elderly populations

Decision-tree inputs shift substantially in older adults and patients with organ dysfunction

Elderly-specific modifications:

Competing risks dominate the tree:

Renal impairment modifies pharmacotherapy branches:

Hepatic impairment changes drug-toxicity utilities:

Polypharmacy introduces interaction nodes:

Step 3 management: When a vignette features an elderly patient with multiple comorbidities, explicitly weigh time to benefit vs. remaining life expectancy — this is the most testable principle in geriatric decision analysis and frequently overrides standard adult guidelines

Board pearl: Deprescribing is a legitimate decision-tree branch, not a failure of care

Special Populations — Pregnancy, Pediatrics, and Equity

— Each chance node may split into maternal AND fetal outcomes

— Utilities for fetal outcomes are ethically and methodologically fraught (often life-years gained, not QALYs)

— Example: anticoagulation in pregnant patient with mechanical valve — warfarin (better maternal valve outcomes, fetal embryopathy 5–10% in first trimester) vs. LMWH (safer for fetus, higher maternal thrombosis if poorly monitored)

— ACE inhibitors, ARBs, warfarin, isotretinoin, valproate, methotrexate carry high-probability fetal-harm branches

— Always substitute before pregnancy when planned, or immediately on diagnosis when unplanned

— Prior unprovoked VTE → antepartum + postpartum LMWH

— Prior provoked VTE → postpartum only

— Tree pivots on recurrence probability vs. bleeding/cost

— Longer time horizon dramatically amplifies QALY gains from preventive interventions (vaccines, lead screening)

— Radiation exposure utility decrement is higher (lifetime cancer risk) — favors ultrasound and MRI branches

— Weight-based dosing introduces error nodes — EHR decision support reduces this branch's probability

— Decision trees built on registry data may underrepresent racial/ethnic minorities, distorting probability estimates

— Race-based equations (eGFR pre-2021, ASCVD Pooled Cohort) are being recalibrated to avoid embedded bias

— Step 3 increasingly tests recognition that structural determinants modify both baseline risk and treatment access

Decision-tree applications in pregnancy must explicitly model two patients

Maternal-fetal trees:

Teratogenicity branches:

VTE prophylaxis decision in pregnancy:

Pediatric considerations:

Health-equity inputs:

Step 3 management: In pregnant patients, never default to a non-pregnant guideline — explicitly check whether the medication, imaging modality, or procedure has a pregnancy-specific branch with different expected value

Board pearl: Shared decision-making in pediatrics is a triadic process (clinician, parent, age-appropriate child) — assent from a 12-year-old is ethically meaningful even when not legally required

Complications and Adverse Outcomes of Decision-Analytic Reasoning

— Omitted branches: missing a relevant outcome (e.g., neglecting contrast-induced nephropathy in a CT-vs-MRI tree)

— Non-exhaustive nodes: probabilities not summing to 1.0

— Overlapping branches: double-counting events that span multiple terminal nodes

— Inappropriate independence assumptions: treating correlated events (stroke and MI, both atherosclerotic) as independent inflates joint probabilities

— Using single trial point estimates without confidence intervals

— Extrapolating efficacy from a trial population to a very different real-world population

— Stale data — guidelines and effect sizes change (e.g., aspirin for primary prevention reversed direction in 2019)

— Patients and clinicians weigh outcomes differently — clinicians often underestimate quality-of-life impact of disability

— Standard gamble vs. time trade-off yield systematically different utilities for the same state

— Anchoring on initial probability estimates

— Availability heuristic inflates recently encountered diagnoses

— Base-rate neglect — ignoring prevalence when interpreting a positive test

— Omission bias — preferring inaction harms over equivalent action harms

— Overtesting cascade: false positives → confirmatory tests → procedural complications

— Therapeutic momentum: empiric treatment that becomes hard to stop

— Financial toxicity: cost as an unmeasured harm utility

Decision trees are powerful but failure-prone — Step 3 tests recognition of their characteristic pitfalls

Modeling errors:

Input errors:

Utility elicitation problems:

Cognitive biases that corrupt clinical use of trees:

Adverse downstream outcomes:

Board pearl: A decision tree that recommends a strategy "dominated" in every sensitivity analysis is conclusive; one whose recommendation flips with plausible input variation requires clinical judgment, not the model

Key distinction: Strong dominance = strategy is cheaper AND more effective; extended (weak) dominance = strategy is dominated by a mixture of others — both eliminate the strategy from consideration

When to Escalate — Beyond the Decision Tree

— Long time horizon (chronic disease over years/decades)

— Recurring events (multiple strokes, repeated hospitalizations)

— Health states that patients move between (CKD stages, NYHA class)

— Example: hepatitis C treatment with DAAs requires modeling fibrosis progression over 20+ years

— Resource constraints matter (OR availability, ICU beds, organ transplant queues)

— Patient interactions affect outcomes (infectious disease transmission)

— Individual-level heterogeneity is critical

— Outcomes can't be reduced to a single utility measure (equity, ethics, patient experience)

— Stakeholder values diverge — guideline panels, formulary committees

— EHR-integrated tools (e.g., MDCalc, UpToDate Pathways) provide validated, updated trees

— Reduces calculation errors and incorporates current evidence

— The "right" branch by expected utility conflicts with patient values, family wishes, or institutional policy

— Goals of care unclear at end of life

— Resource scarcity forces explicit rationing decisions

— Patient explicitly rejects the model-preferred option after informed discussion → honor autonomy

— Rapidly changing clinical status invalidates input probabilities → reassess

— New evidence (recent practice-changing trial) supersedes the tree's data

Recognize when a decision tree is the wrong tool and a different framework is needed

Escalate to a Markov model when:

Escalate to discrete-event simulation when:

Escalate to multi-criteria decision analysis (MCDA) when:

Consult a clinical decision-support system (CDSS) when:

Escalate to ethics consultation when:

Real-time clinical triggers to step back from the tree:

Step 3 management: Recognize that decision trees support — they do not replace — clinical judgment and patient preferences; an answer choice that says "follow the model regardless of patient wishes" is always wrong

CCS pearl: When uncertain, ordering "social work consult," "palliative care consult," or "ethics consult" can be the highest-utility branch in a CCS case

Key Differentials — Related Decision-Analytic Methods

— Cyclic health states with transition probabilities per cycle

— Used for chronic disease (HIV, CKD, dementia)

— Assumes Markov property — future depends only on current state, not history (a limitation often violated and corrected with "tunnel states")

— Individual-level Markov modeling

— Tracks patient history, allowing memory

— Computationally intensive but more realistic

— Outcomes in natural units (life-years, cases prevented)

— ICER expressed as $/life-year

— Outcomes in QALYs

— Most common in modern HTA (NICE, ICER, CADTH)

— Both costs and outcomes monetized

— Controversial because it requires assigning dollar value to life/health

— Short-term affordability for a payer

— Complements but doesn't replace CEA

— Quantifies expected gain from collecting additional data (e.g., funding a new trial)

— Expected value of perfect information (EVPI) sets an upper bound

— Borrowed from finance — values flexibility to defer or modify decisions as information accrues

Decision trees coexist with several adjacent analytic frameworks; distinguishing them is high-yield

Markov models:

Microsimulation:

Cost-effectiveness analysis (CEA):

Cost-utility analysis (CUA):

Cost-benefit analysis (CBA):

Budget impact analysis:

Value of information (VOI) analysis:

Real-options analysis:

Step 3 management: When the vignette asks about lifetime impact of a chronic disease intervention, expect Markov or microsimulation reasoning; when it asks about a single acute decision, the decision tree suffices

Board pearl: All these methods share a common output — expected value — but differ in how they handle time, recurrence, and uncertainty; recognizing which method fits a clinical question is increasingly tested

Key distinction: Sensitivity analysis is not a separate method — it is a component of all of the above

Key Differentials — Non-Quantitative Decision Frameworks

— Synthesize decision-analytic and trial evidence into actionable recommendations

— GRADE methodology grades both quality of evidence and strength of recommendation

— Step 3 expects familiarity with USPSTF (A/B/C/D/I), ACC/AHA (Class I/IIa/IIb/III), and IDSA recommendation schemes

— Convert tree outputs into patient-facing visuals (Option Grids, pictographs)

— Particularly important for preference-sensitive decisions: PSA screening, lumpectomy vs. mastectomy, AF anticoagulation

— Improve knowledge and decisional satisfaction; effect on outcomes mixed

— Centor criteria, Wells score, PERC, CURB-65, HEART score

— Compressed decision trees designed for rapid bedside use

— Each carries a hidden threshold approach beneath it

— Principlism (autonomy, beneficence, non-maleficence, justice)

— When the expected-utility-maximizing branch conflicts with autonomy, autonomy wins (in adults with capacity)

— Patient story and values that resist quantification

— Particularly relevant in end-of-life, chronic pain, mental health decisions

— PDSA cycles, root cause analysis, FMEA (failure modes and effects analysis)

— FMEA is sometimes called the "decision tree of patient safety" — proactively maps how a process can fail

— Highly individualized decisions with unique patient values

— Insufficient or poor-quality data for inputs

— Emergencies where time precludes formal analysis (use protocols instead)

Not every clinical decision should be modeled with a tree; recognize when alternative frameworks apply

Clinical practice guidelines:

Shared decision-making (SDM) aids:

Heuristics and clinical pathways:

Ethical frameworks:

Narrative medicine:

Quality improvement frameworks:

When NOT to use a decision tree:

Board pearl: USPSTF Grade D = "do not provide this service" — actively recommend against; not the same as Grade I (insufficient evidence), where clinical judgment determines action

Key distinction: Guidelines tell you what most patients should do; decision trees and SDM tell you what this patient should do

Secondary Prevention — Integrating Trees into Longitudinal Care

— Disease probabilities evolve (post-MI, post-stroke — risks recalibrate)

— New comorbidities shift utility weights

— Patient preferences mature, especially around end-of-life

— Plan annual or semiannual review of long-term therapies

— High-intensity statin = above treatment threshold for all with established disease

— Add ezetimibe if LDL >70 on max statin; consider PCSK9 inhibitor if LDL still >70 in very-high-risk

— Decision tree increasingly favors aggressive LDL lowering (FOURIER, ODYSSEY trials)

— Provoked, transient risk factor → 3 months

— Unprovoked → indefinite if bleeding risk acceptable

— Cancer-associated → indefinite while cancer active, DOAC preferred (except GI/GU malignancy)

— HERDOO2 rule helps women below threshold for indefinite therapy

— Aspirin + P2Y12 inhibitor (DAPT typically 12 months)

— High-intensity statin

— Beta-blocker (especially if reduced EF or recent MI)

— ACE inhibitor/ARB if EF <40%, HTN, DM, or CKD

— Aldosterone antagonist if EF <40% with HF or DM

— A1c goals individualized (general 7%, relaxed to 8% in frail/elderly/limited life expectancy)

— SGLT2 inhibitor or GLP-1 RA regardless of A1c in ASCVD/CKD/HF

Decision trees inform not just initial choices but ongoing management over months and years

Periodic reassessment — inputs change:

Statin therapy after ASCVD event:

Anticoagulation duration after VTE:

Post-MI medication bundle (each justified by its own NNT-derived branch):

Diabetes secondary prevention:

Step 3 management: Discharge medication lists should be complete and individualized — each medication justified by a decision-tree branch; omitting evidence-based therapy at discharge is a tested patient safety failure

Board pearl: Adherence is itself a chance node — a 100%-effective medication taken 50% of the time yields 50% benefit

Follow-Up, Monitoring, and Counseling

— Post-MI: cardiology in 2–6 weeks, primary care in 7–14 days

— New anticoagulation: 1–2 weeks for adherence and bleeding check, then 3-month intervals

— New antihypertensive: 4 weeks for response, then 3–6 months once controlled

— Diabetes: A1c q3 months if uncontrolled, q6 months if stable

— Chronic kidney disease: depends on stage and albuminuria — KDIGO heat-map drives frequency

— Each lab/imaging follow-up is a decision: continue, escalate, de-escalate

— Example: warfarin INR — out of range → dose adjustment branch; in range → continue

— Teach-back method confirms understanding (a quality measure)

— Address health literacy (5th–8th grade reading level for written materials)

— Disclose probabilities in natural frequencies ("3 out of 100 patients") rather than percentages alone — improves comprehension

— Visual aids (pictographs, icon arrays) outperform verbal disclosure

— Cardiac rehab post-MI: Class I recommendation, ~25% mortality reduction, dramatically underutilized

— Pulmonary rehab in COPD with GOLD B+ disease

— Stroke rehab — inpatient vs. SNF vs. outpatient determined by functional status

— Medication reconciliation at every transition

— Post-discharge phone call within 48 hours reduces readmission

— Follow-up appointment within 7 days for high-risk diagnoses (HF, COPD exacerbation)

Follow-up cadence is the time axis of a decision tree — too long misses events, too short generates false alarms and cost

Evidence-based follow-up intervals:

Monitoring parameters as chance nodes:

Patient counseling components:

Rehabilitation as a decision branch:

Transitions of care generate the highest-risk branches:

Step 3 management: Every discharge order set should include specific follow-up timing, pending lab/test results, warning signs to return for, and confirmed medication list — these are the four pillars of safe transition and frequently tested

Board pearl: The "no-show" branch of follow-up has measurable consequences — proactive outreach is a quality intervention

Ethical, Legal, and Patient Safety Considerations

— Disclosure must include alternatives, including no treatment

— Numerical risks should be communicated in absolute terms with confidence intervals, not just relative risks

— Patient with capacity may refuse the expected-utility-maximizing option — this is autonomy, not non-adherence

— Four components: understanding, appreciation, reasoning, expressing a choice

— Capacity is decision-specific — a patient may have capacity for some choices but not others

— Disagreement with clinician does not imply incapacity

— Cost-effectiveness thresholds raise equity concerns — strategies cost-effective in average populations may not serve marginalized groups

— Step 3 tests recognition that denying care based purely on cost without transparent process is ethically problematic

— Decision-tree authors with industry funding may inflate efficacy or omit harms

— Disclosure required in any guideline or analysis

— Suspected child/elder abuse, certain infectious diseases, gunshot wounds, impaired drivers (state-dependent), Tarasoff-style duty to warn

— These bypass standard confidentiality — a hard-coded branch

— Every transition of care (admission, transfer, discharge) is a high-risk node

— Studies show ~50% of patients have at least one medication discrepancy at discharge without formal reconciliation

— Pharmacist-led reconciliation reduces adverse drug events by ~30%

— This is a concrete Step 3 patient safety priority rooted in transition-of-care risk

— Ethical duty to disclose harm-causing errors transparently

— Apology laws in most states protect compassionate disclosure from use as evidence of liability

Decision trees raise distinct ethical issues that Step 3 explicitly tests

Informed consent and decision aids:

Capacity assessment:

Justice and resource allocation:

Conflicts of interest:

Mandatory reporting branches:

Patient safety — the medication reconciliation case:

Disclosure of medical errors:

Board pearl: "Do no harm" includes the harm of overuse — testing and treating low-probability disease causes net harm even when individual tests seem benign

High-Yield Associations and Rapid-Fire Facts

Branches must sum to 1.0 at every chance node

Expected value = Σ (probability × outcome value), folded back right to left

QALY = years × utility weight (0=death, 1=perfect health)

ICER = ΔCost / ΔEffectiveness, threshold ~$50K–$150K/QALY in US

Sensitivity analysis tests robustness; tornado diagram ranks input influence

Testing threshold below which testing harms exceed benefits

Treatment threshold above which empiric treatment beats testing

Pauker-Kassirer three-zone model — foundational

Likelihood ratios drive Bayesian probability revision: LR >10 strong positive, <0.1 strong negative

Pretest odds × LR = posttest odds

Number needed to treat (NNT) = 1/ARR

Number needed to harm (NNH) = 1/ARI

Markov model for chronic, recurrent disease — has the Markov property (memoryless)

Decision tree for single-stage, time-limited decision

Dominance: strategy cheaper AND more effective — eliminates competitor

Time to benefit must be shorter than remaining life expectancy for prevention to make sense

Discount rate 3%/year for costs and benefits in US analyses

Standard gamble and time trade-off are utility elicitation methods

Lead-time bias and length-time bias distort screening evaluations

Berkson bias affects hospital-based decision analyses

CHEERS checklist = reporting standard for economic evaluations

GRADE = recommendation strength framework integrating evidence + values

USPSTF grades: A/B (recommend), C (selective), D (recommend against), I (insufficient)

Shared decision-making essential for preference-sensitive conditions

Heuristics (Centor, Wells, HEART, PERC) are compressed decision trees

EVPI = expected value of perfect information — upper bound on study value

One-way, two-way, probabilistic sensitivity analyses — progressively more comprehensive

Patient autonomy trumps expected utility when in conflict

Board pearl: Decision quality ≠ outcome quality — a good decision can yield a bad outcome; this distinction underpins both quality improvement and malpractice analysis

Board Question Stem Patterns

— "A 65-year-old man with chest pain has 40% pretest probability of CAD. Stress test sensitivity 80%, specificity 75%. Next step?"

— Recognize the between-thresholds zone → test, don't treat empirically, don't dismiss

— "Disease prevalence 10%, test sensitivity 90%, specificity 80%. PPV of a positive test?"

— Build 2×2 table with 1,000 patients; PPV ≈ 33% — illustrates base-rate neglect trap

— Two treatments, each with probability/outcome pairs

— Calculate Σ(p × outcome) for each; pick higher EV — beware "highest possible outcome" distractor

— Strategy A: $100K, 5 QALY; Strategy B: $150K, 6 QALY

— ICER = $50K/QALY → cost-effective at typical US threshold

— Strategy that is both cheaper AND more effective → dominant, always preferred

— Strategy dominated by a mixture → extended dominance, also eliminated

— "Recommendation unchanged across plausible range of all inputs" → robust, adopt

— "Recommendation flips with small change in stroke risk" → fragile, individualize

— Patient with capacity refuses expected-value-best option after full disclosure

— Correct answer: honor refusal, document, offer alternatives

— Discharge scenario with missing follow-up or unreconciled medications

— Correct answer: medication reconciliation, scheduled follow-up, clear return precautions

— Elderly patient with limited life expectancy considering screening

— Correct answer: weigh lead time vs. life expectancy — often defer

Pattern 1 — The threshold question:

Pattern 2 — The Bayesian update:

Pattern 3 — The expected-value calculation:

Pattern 4 — The ICER question:

Pattern 5 — The dominance question:

Pattern 6 — The sensitivity analysis question:

Pattern 7 — The autonomy override:

Pattern 8 — The transition-of-care safety question:

Pattern 9 — The time-to-benefit question:

Board pearl: When a question gives you specific numerical probabilities and outcome values, the answer is almost always calculate, don't intuit — even rough mental math will identify the dominant strategy

One-Line Recap

High-yield recap bullets:

Decision trees translate clinical uncertainty into structured, quantifiable comparisons by mapping decision and chance nodes to outcomes weighted by probability and utility — guiding evidence-based, patient-centered choices through expected-value calculation, threshold analysis, and sensitivity testing.

Three-zone threshold model (Pauker-Kassirer) is the workhorse: below testing threshold → don't test; between → test then treat; above treatment threshold → treat empirically — and the most testable single framework on Step 3 biostatistics

Fold back from right to left, choosing the maximum expected value at each decision node and averaging out at each chance node; the branch with highest expected utility, NOT the highest possible outcome, wins — confusing these two is the single most common exam trap

Sensitivity analysis (one-way, two-way, probabilistic) determines whether a recommendation is robust; tornado diagrams identify which inputs matter most; recommendations stable across plausible input ranges are guideline-worthy

Patient autonomy and shared decision-making override expected-utility maximization when a patient with capacity, after full informed disclosure including absolute risks and alternatives, prefers a different branch — and transition-of-care safety (medication reconciliation, timely follow-up, clear return precautions, communicated pending results) is the concrete operational embodiment of decision-analytic principles at the bedside, repeatedly tested on Step 3 CCS and MCQ alike

Board pearl: A good decision is one made correctly given available information — not one guaranteed to yield a good outcome; separating decision quality from outcome quality is the philosophical core of clinical decision analysis