Guide to Common Measures (Data type, Clinimetrics) in Clinical Research Using the DEPTh Model
🔍 Why Clinical Metrics Matter (Beyond “Significance”)
Imagine you’re comparing two treatments in a heart failure unit. Or evaluating a new diagnostic AI in diabetic retinopathy. Or modeling survival in glioblastoma.
In each case, you’ll eventually report clinical measures—but if you don’t start from a clear DEPTh challenge, you risk matching the wrong metric to the wrong question.
🩺 Key Principle: A measure only matters if it’s aligned to the clinical decision you're informing.
This is where the DEPTh model reframes your approach:
| DEPTh Challenge | Core Clinical Question |
| Diagnosis | Is test X accurate for detecting disease Y? |
| Etiognosis | Is exposure X a cause or risk factor for outcome Y? |
| Prognosis | What’s the likely outcome for a patient with Z? |
| Therapeutic | Does treatment X improve patient outcome Y? |
| Methodologic | How should we best measure/model/analyze? |
🧪 Clinimetrics = Metrics with a Mission
Clinimetrics combines clinical logic with measurement theory. You aren’t just calculating numbers—you’re quantifying questions.
To build metrics that matter, first structure your study using the Design Triad:
- Object Design: DEPTh challenge
- Method Design: Domain, base, determinant (X), endpoint (Y), calendar time
- Analysis Design: Causal (explanatory) vs Predictive (risk modeling)
🧠 Start With the DEPTh Type
Let’s now walk through each DEPTh category, replacing old examples with fresh, original, and clinically realistic cases.
💊 Therapeutic Research: Comparing Interventions
🎯 Research Question
Does continuous positive airway pressure (CPAP) reduce morning systolic blood pressure more than high-flow nasal cannula (HFNC) in patients with OSA?
🔬 Study Design
- RCT with binary X: CPAP (1), HFNC (0)
- Continuous Y: Morning SBP (mmHg)
🧮 Metrics
| Outcome Type | Preferred Metric | Interpretation |
| Continuous | Mean difference / Mean ratio | Average BP reduction |
| Binary | Risk ratio, Odds ratio, NNT | Proportion with controlled BP |
| Time-to-event | Hazard ratio, Rate ratio | Time to normalization |
Example result: Mean SBP after CPAP: 124.2 mmHg Mean SBP after HFNC: 133.4 mmHg Mean Difference = −9.2 mmHg 95% CI: (−13.1, −5.3) → Clinically meaningful reduction
🔍 Secret Insight: Use MCID to interpret whether this 9.2 mmHg drop crosses the “minimal important” threshold for decision-making.
⚠️ Etiognostic Research: Identifying Risk Factors
🎯 Research Question
Does high sodium intake increase the risk of frequent nocturia in older adults?
🔬 Study Design
- Prospective cohort
- Binary X: High sodium diet (1), normal sodium (0)
- Binary Y: Nocturia ≥2 times/night (1), otherwise (0)
🧮 Metrics
| Metric | Use Case |
| Risk Ratio | Incidence in exposed vs unexposed |
| Odds Ratio | Use in case-control or rare outcomes |
| Hazard Ratio | Use if timing to event matters |
Example result:
- High sodium group risk = 30%
- Normal sodium group risk = 12% Risk Ratio = 2.5 Risk Difference = 18%
🛠 Use DAGs to rule out confounding by other dietary or renal variables before interpreting causality.
⏳ Prognostic Research: Predicting Outcomes
🎯 Research Question
Among stroke survivors aged >70, what predicts functional independence at 6 months?
🔬 Study Design
- Prospective cohort of post-stroke patients
- X = Age, NIHSS score, rehab intensity (continuous/categorical)
- Y = Modified Rankin Scale (mRS ≤2 = good outcome)
🧮 Metrics
| Metric | Purpose |
| Odds Ratio | Risk prediction modeling (logistic regression) |
| AUROC | Discrimination of the model |
| Calibration Plot | Agreement between predicted and actual rates |
| Kaplan-Meier | For time-to-functional-recovery modeling |
Example result:
- AUROC = 0.82 (very good discrimination)
- NIHSS ≥10 associated with OR = 0.3 for good outcome
🔍 Secret Insight: In prognostic modeling, calibration is just as important as discrimination—models must be right on average and accurate across risk ranges.
🧪 Diagnostic Research: Measuring Test Accuracy
🎯 Research Question
Can a new retinal AI scan accurately detect early-stage diabetic retinopathy?
🔬 Study Design
- Cross-sectional study
- Binary X = AI prediction: DR detected (1), not detected (0)
- Binary Y = Ophthalmologist-confirmed DR (gold standard)
🧮 Metrics
| Metric | Interpretation |
| Sensitivity | Ability to detect actual DR |
| Specificity | Ability to exclude non-DR |
| LR+ / LR– | Change in odds after test result |
| AUROC | Global accuracy |
| PPV / NPV | Population-specific predictive values |
Example result:
- Sensitivity = 88%
- Specificity = 91%
- AUROC = 0.94
⚠️ Predictive values vary by prevalence. Always report PPV/NPV with context.
🔗 Aligning Measures to Study Logic
Use this master table to pair your X-Y type with the right metric:
| X Type | Y Type | Metric Options |
| Binary | Binary | RR, OR, RD, NNT |
| Binary | Continuous | Mean difference, ratio |
| Binary | Time-to-event | Hazard ratio, rate ratio |
| Continuous | Binary | Regression coefficient, OR |
| Continuous | Time-to-event | Cox proportional hazard model |
✅ Key Takeaways
- Always start from the DEPTh classification—it defines everything.
- Use the Design Triad to structure your study: object, method, analysis.
- Match your metrics to both data type and research goal (causal vs predictive).
- Don’t interpret significance without effect size + 95% CI.
- Use MCID and clinical impact framing to go beyond p-values.