Literature intelligence

Browse PROBAST-gated evidence, relevance scores, and how each article class feeds the DOC knee-OA model stack - TKR risk, WOMAC & NRS outcomes, and conservative treatment estimates.

← Main clinical app Performance metrics → Risk model charts →

Model stack — evidence mapping

TKR necessity (4-year)

Random Forest AUC 0.92
Role of literature
Optional literature-calibrated probability layer (Platt-style) aligns risk percentages with published cohorts while preserving discrimination.
Artifacts
random_forest_best.pkl · random_forest_literature_calibrated_base.pkl
Eligible papers
From literature.db: relevance ≥40, PROBAST domains 1–4 Low or Moderate (no High).

Post-TKR outcome - WOMAC

Pooled regressor OAI + Khatib
Role of literature
Outcome calibration (optional) nudges predicted improvement toward meta-analytic / registry means; core model remains OAI+Khatib–trained.
Artifacts
outcome_rf_regressor_pooled.pkl · womac_outcome_calibration.pkl (if present)
Symptom scale
WOMAC total 0–96; success metrics derived from predicted improvement bands.

Post-TKR outcome - NRS / VAS

Bergman-trained N=4,426
Role of literature
Native NRS scale for European workflows; same clinical inputs mapped to Bergman feature space.
Artifacts
nrs_outcome_model.pkl + scaler / imputer / feature names
Calibration
Optional nrs_outcome_calibration.pkl when literature alignment is enabled in app.

Conservative pathways

Exploratory Shared decision
Sub-models
Surgery avoidance (logistic), NSAID / corticosteroid / HA response (RF-style), nutrition pathway when NRS scale is active.
Literature blend
conservative_literature_calibration.json - conservative-only evidence, separate from TKR core EPV.
Use
Context for trial of non-operative care vs expected surgical trajectory — not a replacement for RCT-level evidence.

PROBAST gate (prediction models)

Articles in the model pool must meet relevance ≥ 40 and have no High risk domain. Up to one Moderate domain is allowed with documentation - consistent with your top-tier quality bar.

D1 Participants D2 Predictors D3 Outcome D4 Analysis
Article quality rubric (expand)

Usable for model weighting: relevance ≥40 + all four PROBAST domains Low or Moderate (no High).

  • Domain 1 — Participants: Clear eligibility, representative knee-OA population.
  • Domain 2 — Predictors: Predictors defined a priori, measured before outcome.
  • Domain 3 — Outcome: Validated outcome definition (e.g. WOMAC, TKR).
  • Domain 4 — Analysis: Adequate EPV, validation, missing data handled.
-
Total articles
-
Open access
-
Paywalled
-
Model-eligible (PROBAST)
Title Journal Score Access PMID