Prognosis Research: Frameworks, Metrics, and Clinical Integration

Mayta
May 16
3 min read

🚦 What Clinical Question Are We Tackling?

DEPTh classification: Prognosis

Framing: "Given a patient’s diagnosis or clinical state, what is likely to happen next?"

This framing distinguishes prognostic inquiry from causal or diagnostic aims. Prognosis tells the story after diagnosis, not what caused the condition.

📐 Theoretical Foundation: Prognosis ≠ Cause

As clarified in the Prognostic_CECS and Causal Inference Guide, prognosis research does not test causality. Rather, it characterizes future outcomes in diagnosed patients without making assumptions about what caused them.

Occurrence Equation for Prognosis:

🧭 The PROGRESS Framework – 4 Pillars

PROGRESS I: Overall Prognosis Research

Goal: Describe typical outcome patterns.

Design: Descriptive cohort.
Metrics: Median survival, cumulative incidence.
Clinical Utility: Benchmarks, public health surveillance.

📌 Example: What is the 2-year survival for IPF patients in Northeast Thailand?

🔍 Secret Insight: Median ≠ mean ≠ message! Median survival hides long tails—never use it alone for patient counseling.

PROGRESS II: Prognostic Factor Research

Goal: Identify variables associated with outcomes.

Design: Inception cohort with longitudinal follow-up.
Metrics: Hazard Ratios (HR), but prioritize absolute risks (e.g., 30-day mortality).
Impact: Guides triage, follow-up intensity, and patient counseling.

📌 Example: Does elevated troponin predict 30-day mortality in dengue shock?

🔍 Secret Insight: Prognostic ≠ Predictive. Just because a factor signals bad outcome doesn’t mean it tells who benefits from treatment.

PROGRESS III: Prognostic Model Research (CPRs)

Goal: Predict individual risk by combining multiple factors.

Design: Cohort with well-defined predictors and outcomes.
Process: 9-phase roadmap (from Clinical Prediction Models_CECS) includes:
1. Define need
2. Fix prediction point
3. Choose predictors
4. Handle missing data (MI)
5. Model derivation (Cox/logistic)
6. Evaluate (AUROC, calibration)
7. Validate (internal & external)
8. Assess clinical impact

📌 Example: Build and validate a model for predicting 6-month mortality in cirrhotic patients with GI bleeding.

🔍 Secret Insight: Resist the urge to build anew—adapt existing models unless you can prove a new context or unmet need (use the "🚦traffic light" logic).

PROGRESS IV: Stratified Medicine Research

Goal: Determine who benefits most/least from a therapy.

Design: Subgroup or interaction analysis in RCTs or target trial emulation.
Analytic Tools: Effect modification analysis; interaction terms.

📌 Example: Does anemia modify the effect of furosemide in ICU patients with heart failure?

🔍 Secret Insight: Interaction ≠ association. Always test if subgroup effects are statistically significant and clinically plausible.

⚙️ Method Design: Building the Right Frame

1. Study Domain

"Patients to be prognosticated"—clearly define who and when.

📌 Example: “Adults with ischemic stroke admitted within 24h to Songklanagarind Hospital (2016–2020).”

2. Study Base

Clarify whether it's prospective registry, retrospective EHR, or RCT-derived.

3. Determinants (Predictors)

Include only those measurable at or before the inception point. Ensure clinical interpretability.

4. Outcome Definitions

Binary: e.g., 30-day mortality
Time-to-event: e.g., time to readmission
Continuous: e.g., LOS

📌 Pro Tip: Define the time horizon explicitly—72-hour vs 6-month outcomes serve different clinical goals.

🧮 Analytic Design: Predict or Explain?

Objective	Tools	Missteps
Prediction	Risk models, AUROC, calibration	Overfitting, optimism bias
Explanation	DAGs, confounder adjustment	Collider/mediator confusion

From Causal Inference Guide: Prediction ≠ explanation. Don’t “adjust” blindly—use DAG-first logic only for causal inference.

🧪 Key Metrics & Tools

PROGRESS Type	Key Tools	Metrics
I	Kaplan-Meier	Median survival, Cumulative Incidence
II	Cox regression, logistic regression	HR, OR, Absolute Risk
III	Multivariable regression, ML	AUROC, Calibration, NRI, DCA
IV	Interaction models, stratified HRs	Effect modification, Subgroup curves

🧬 Clinical Integration

Prognostic insights shape:

ICU vs ward admission
End-of-life discussions
Palliative vs aggressive treatment
Digital decision-support tools (EMRs)

📌 Implementation note: Clinical Prediction Models (CPMs) are only useful if trusted and usable at the bedside. Test impact via stepped-wedge trials or cluster-RCTs.

✅ Key Takeaways

Prognosis research answers “What happens next?”—not “Why?”
Always define a clear point of prediction—the moment all predictors are observable.
Anchor your design in PROGRESS I–IV logic.
Distinguish absolute vs relative risk.
CPMs must be justified, well-calibrated, and validated before clinical use.