⭐ 高 yield 摘要
Dual-Process Theory
| System 1 |
快直覺 |
Pattern matching |
Premature closure |
| System 2 |
慢分析 |
Hypothetico-deductive |
Decision paralysis |
4 個 Heuristics(必認)
| Representativeness |
「像 X 病」 |
忽略 base rate |
| Availability |
「最近 case」 |
受 media / recent encounters 影響 |
| Anchoring |
「第一印象」 |
不依新證據調整 |
| Occam’s razor |
「single dx」 |
老人多重共病無解 |
Hickam’s Dictum
- 「Patients can have as many diseases as they damn well please」
- 適用:老人、複雜共病
- 對立 Occam’s razor
Diagnostic Imperatives
- 「rare but catastrophic if missed」必排除
- 例:胸痛 → 排 aortic dissection;頭痛 → 排 SAH;發燒 → 排 meningitis
2x2 Table 與檢查特性
Disease+ Disease-
Test+ TP FP
Test- FN TN
| Sn |
TP/(TP+FN) |
❌ 不受 |
| Sp |
TN/(TN+FP) |
❌ 不受 |
| PPV |
TP/(TP+FP) |
✅ 強烈 |
| NPV |
TN/(TN+FN) |
✅ 強烈 |
Likelihood Ratios 速查
| > 10 |
強強確診 |
| 5-10 |
中等 |
| 2-5 |
弱 |
| ≈ 1 |
沒用 |
| < 0.1 |
強強排除 |
| 0.1-0.5 |
中等 |
| > 0.5 |
弱 |
Bayes’ Theorem 實作
- Pre-test prob → odds:
odds = p/(1-p)
- Post-test odds = pre-test odds × LR
- Post-test prob = odds/(1+odds)
Wells Criteria for PE(表 4-2,必背)
| Clinical signs DVT |
3 |
| PE 是 most likely dx |
3 |
| HR > 100 |
1.5 |
| Immobilization ≥3 day or surgery in 4 wk |
1.5 |
| Previous DVT/PE |
1.5 |
| Hemoptysis |
1 |
| Malignancy |
1 |
| > 6 |
High |
| 2-6 |
Intermediate |
| < 2 |
Low |
EBM 4 步驟
- Ask:PICO 格式(Population/Intervention/Comparator/Outcome)
- Acquire:systematic review > RCT > cohort > case series
- Appraise:critically appraise(GRADE 等)
- Apply:integrate with clinical judgment + patient preference
Evidence Hierarchy
Systematic review of RCTs (highest)
↓
Single RCT
↓
Cohort study
↓
Case-control study
↓
Case series
↓
Expert opinion (lowest)
Practice Variation 三類影響因子
- Practice Style:個人訓練、經驗、subspecialty 影響(cardiologist vs internist 不同)
- Practice Setting:work system、tech、organization、physical env
- Payment System:fee-for-service vs capitation vs P4P 影響
Diagnostic Errors(系統議題)
- 預估每人一生至少 1 次 diagnostic error
- 解法:EHR Decision Support、checklists、reminder system
AI 三類
- Neural networks
- Machine learning(含 deep learning)— image / signal interpretation
- Generative AI(LLM)— note-taking、early decision support
22E 立場:「Not abdicate decision-making entirely to algorithm」