Skip to contents

Bayesian Decision Curve Analysis - Advanced Clinical Decision Making

Comprehensive guide to evaluating clinical utility of prediction models and diagnostic tests

ClinicoPath

2025-07-13

Source: vignettes/meddecide-07-bayesian-decision-curve-analysis.Rmd
meddecide-07-bayesian-decision-curve-analysis.Rmd

Introduction to Decision Curve Analysis

Decision Curve Analysis (DCA) is a powerful method for evaluating the clinical utility of prediction models and diagnostic tests. Unlike traditional performance metrics (sensitivity, specificity, AUC), DCA directly addresses the question: “Would using this model lead to better clinical decisions?”

What Makes This Analysis Special?

Clinical Utility Focus

  • Moves beyond discrimination metrics to real-world decision making
  • Incorporates the relative harm of false positives vs false negatives
  • Provides threshold-specific guidance for clinical decisions
  • Quantifies net clinical benefit across decision thresholds

Bayesian Enhancement

  • Uncertainty quantification through posterior distributions
  • Credible intervals for net benefit estimates
  • Probability of superiority between competing strategies
  • Expected Value of Perfect Information (EVPI) for research prioritization

When to Use Decision Curve Analysis

Perfect for: - Evaluating prediction models for clinical implementation - Comparing diagnostic tests with different cost-benefit profiles - Determining optimal decision thresholds for interventions - Assessing value of additional biomarkers or clinical variables - Health economic evaluation of diagnostic strategies

Examples: - Should we implement a cancer screening model in practice? - At what probability threshold should we recommend treatment? - Is an expensive biomarker worth the additional cost? - Which prediction model provides the most clinical benefit?

Theoretical Foundation

The Net Benefit Framework

Decision Curve Analysis is based on the net benefit metric:

NB=Sensitivity×Prevalence−False Positive Rate×(1−Prevalence)×pt1−ptNB = \text{Sensitivity} \times \text{Prevalence} - \text{False Positive Rate} \times (1-\text{Prevalence}) \times \frac{p_t}{1-p_t}

Where: - Sensitivity: True positive rate of the test/model - Prevalence: Disease prevalence in the population - ptp_t: Decision threshold (probability at which you would intervene) - pt1−pt\frac{p_t}{1-p_t}: Odds ratio representing relative harm of false positives

Understanding Decision Thresholds

The decision threshold represents the probability at which a clinician would recommend intervention:

  • Low thresholds (1-5%): Used when missing disease is very harmful (e.g., cancer screening)
  • Medium thresholds (10-20%): Balanced scenarios (e.g., cardiac interventions)
  • High thresholds (30-50%): Used when intervention has significant risks (e.g., major surgery)

Bayesian vs Frequentist Approaches

Aspect Frequentist DCA Bayesian DCA
Uncertainty Bootstrap confidence intervals Posterior credible intervals
Interpretation Long-run frequency properties Probability statements about parameters
Prior Information Not incorporated Can incorporate expert knowledge
Computational Bootstrap resampling MCMC/analytical posteriors
Decision Making Point estimates with CI Full uncertainty distribution

Getting Started

Load Required Libraries 

library(ClinicoPath)
library(dplyr)
library(ggplot2)

# Use the histopathology dataset for examples
data("histopathology")
mydata <- histopathology

# Display basic dataset information
cat("Dataset dimensions:", nrow(mydata), "rows ×", ncol(mydata), "columns\n")
## Dataset dimensions: 250 rows × 38 columns
cat("Outcome variable 'Death':", table(mydata$Death), "\n")
## Outcome variable 'Death': 169 80
cat("Available predictors:", paste(c("Age", "Grade", "TStage", "MeasurementA", "MeasurementB"), collapse = ", "), "\n")
## Available predictors: Age, Grade, TStage, MeasurementA, MeasurementB

Basic Workflow Overview

The Bayesian DCA workflow follows these steps:

  1. Define Binary Outcome: Select outcome variable (0/1 or factor)
  2. Select Predictors: Choose models/tests to evaluate (continuous probabilities or binary results)
  3. Set Thresholds: Define clinically relevant decision threshold range
  4. Choose Analysis Type: Bayesian (with uncertainty) or Frequentist (with bootstrap CI)
  5. Interpret Results: Evaluate net benefit, probability of superiority, and EVPI

Core Analysis Examples

Example 1: Basic Bayesian Decision Curve Analysis

Let’s evaluate whether age and tumor measurements can guide treatment decisions for cancer patients.

# Basic Bayesian DCA comparing age and measurements as predictors
bayesdca(
  data = mydata,
  outcomes = "Death",
  predictors = c("Age", "MeasurementA", "MeasurementB"),
  thresholdMin = 0.05,     # 5% minimum threshold
  thresholdMax = 0.40,     # 40% maximum threshold
  thresholdPoints = 30,    # 30 evaluation points
  bayesianAnalysis = TRUE,
  nDraws = 2000,          # Number of posterior draws
  priorStrength = 2       # Weak prior
)

What you get: - Net benefit curves with credible intervals - Probability of each strategy being optimal - Strategy comparison across thresholds - Comprehensive statistical tables

Example 2: Frequentist Analysis with Bootstrap

For comparison, run the same analysis using frequentist methods:

# Frequentist DCA with bootstrap confidence intervals
bayesdca(
  data = mydata,
  outcomes = "Death", 
  predictors = c("Age", "MeasurementA", "MeasurementB"),
  thresholdMin = 0.05,
  thresholdMax = 0.40,
  thresholdPoints = 30,
  bayesianAnalysis = FALSE,  # Use frequentist approach
  bootstrapCI = TRUE,
  bootstrapReps = 2000      # Bootstrap replications
)

Differences: - Bootstrap confidence intervals instead of credible intervals - No probability statements about superiority - No EVPI calculations available

Example 3: Binary Diagnostic Test Evaluation

Evaluate binary test results (e.g., imaging findings, genetic markers):

# Convert measurements to binary tests for demonstration
mydata_binary <- mydata %>%
  mutate(
    HighMeasurementA = ifelse(MeasurementA > median(MeasurementA, na.rm = TRUE), 1, 0),
    PositiveGrade = ifelse(Grade %in% c("High"), 1, 0)
  )

# Analyze binary diagnostic tests
bayesdca(
  data = mydata_binary,
  outcomes = "Death",
  predictors = c("HighMeasurementA", "PositiveGrade"),
  thresholdMin = 0.10,
  thresholdMax = 0.50,
  bayesianAnalysis = TRUE,
  nDraws = 2000
)

Binary Test Features: - Tests are applied directly without threshold cutoffs - Results show performance at different decision thresholds - Useful for imaging findings, genetic tests, lab markers

Example 4: External Prevalence Adjustment

When your study population differs from the target population:

# Use external prevalence data (e.g., from registry)
bayesdca(
  data = mydata,
  outcomes = "Death",
  predictors = c("Age", "MeasurementA"),
  thresholdMin = 0.05,
  thresholdMax = 0.35,
  useExternalPrevalence = TRUE,
  externalCases = 150,      # Known cases in target population
  externalTotal = 1000,     # Total target population
  bayesianAnalysis = TRUE
)

When to Use: - Study cohort has different characteristics than target population - Registry data provides better prevalence estimates - Multi-site validation with varying prevalence

Advanced Features

Expected Value of Perfect Information (EVPI)

EVPI quantifies the maximum value of conducting additional research to reduce uncertainty:

# Full analysis with EVPI calculation
bayesdca(
  data = mydata,
  outcomes = "Death",
  predictors = c("Age", "MeasurementA", "MeasurementB", "TStage"),
  thresholdMin = 0.05,
  thresholdMax = 0.40,
  bayesianAnalysis = TRUE,
  calculateEVPI = TRUE,    # Enable EVPI calculation
  nDraws = 3000           # More draws for stable EVPI
)

EVPI Interpretation: - High EVPI: Additional research would be valuable - Low EVPI: Current evidence is sufficient for decision making - Peak EVPI: Threshold ranges where research is most valuable - Zero EVPI: Perfect certainty about optimal strategy

Direction Control for Predictors

Control how continuous predictors are interpreted:

# For biomarkers where LOWER values indicate risk
bayesdca(
  data = mydata,
  outcomes = "Death",
  predictors = "MeasurementB",
  directionIndicator = "<=",  # Lower values = positive prediction
  thresholdMin = 0.10,
  thresholdMax = 0.40,
  bayesianAnalysis = TRUE
)

Direction Options: - “>=”: Higher values indicate positive outcome (default) - “<=”: Lower values indicate positive outcome

Prior Strength Sensitivity Analysis

Explore the impact of different prior specifications:

# Weak prior (default)
bayesdca(
  data = mydata,
  outcomes = "Death",
  predictors = "Age",
  priorStrength = 2,      # Weak prior
  bayesianAnalysis = TRUE
)

# Strong prior (more conservative)
bayesdca(
  data = mydata,
  outcomes = "Death", 
  predictors = "Age",
  priorStrength = 10,     # Strong prior
  bayesianAnalysis = TRUE
)

Prior Strength Guide: - 0.5-2: Very weak prior (data-driven) - 2-5: Weak prior (default range) - 5-10: Moderate prior (conservative) - >10: Strong prior (very conservative)

Comprehensive Clinical Example

Scenario: Cancer Treatment Decision Model

Let’s work through a complete analysis for a cancer treatment decision model:

# Comprehensive cancer treatment decision analysis
cancer_dca <- bayesdca(
  data = mydata,
  
  # Core variables
  outcomes = "Death",
  predictors = c("Age", "TStage", "MeasurementA", "MeasurementB"),
  
  # Threshold settings (5% to 40% - typical for cancer treatment)
  thresholdMin = 0.05,
  thresholdMax = 0.40,
  thresholdPoints = 35,
  
  # Bayesian analysis with EVPI
  bayesianAnalysis = TRUE,
  nDraws = 3000,
  priorStrength = 2,
  calculateEVPI = TRUE,
  
  # External prevalence (if applicable)
  useExternalPrevalence = FALSE,
  
  # Classification direction
  directionIndicator = ">="
)

Interpreting the Results

1. Main Decision Curves: - Shows net benefit for each strategy across thresholds - Curves above “treat none” and “treat all” indicate clinical utility - Higher curves = better clinical performance

2. Net Benefit Differences: - Shows improvement over default strategies (treat all/none) - Positive values indicate beneficial use of the model - Confidence bands show uncertainty in benefit

3. Probability of Superiority: - Bayesian probability each strategy is optimal - Helps identify threshold ranges where models are most reliable - Values >80% suggest strong evidence for superiority

4. Expected Value of Perfect Information: - Maximum benefit of reducing uncertainty through research - High EVPI suggests value in additional studies - Can inform research prioritization and funding decisions

Parameter Reference Guide

Core Parameters 

bayesdca(
  data = mydata,                    # Required: Data frame
  
  # Required variables
  outcomes = "Death",               # Binary outcome (0/1 or factor)
  outcomePos = "Yes",              # Positive level (if factor)
  predictors = c("Age", "Test1"),   # Models/tests to evaluate
  
  # Threshold settings
  thresholdMin = 0.01,             # Minimum threshold (0.1-50%)
  thresholdMax = 0.50,             # Maximum threshold (1-99%)
  thresholdPoints = 50,            # Number of evaluation points
  
  # Analysis type
  bayesianAnalysis = TRUE,         # TRUE=Bayesian, FALSE=Frequentist
  
  # Bayesian settings
  nDraws = 2000,                   # Posterior draws (500-10000)
  priorStrength = 2,               # Prior strength (0.1-10)
  calculateEVPI = FALSE,           # Enable EVPI calculation
  
  # Frequentist settings  
  bootstrapCI = TRUE,              # Bootstrap confidence intervals
  bootstrapReps = 2000,            # Bootstrap replications
  
  # External prevalence
  useExternalPrevalence = FALSE,   # Use external prevalence data
  externalCases = 100,             # Cases in external data
  externalTotal = 500,             # Total in external data
  
  # Technical settings
  directionIndicator = ">="        # Prediction direction
)

Parameter Selection Guidelines

Threshold Range Selection

Clinical Scenario Suggested Range Rationale
Cancer Screening 1-10% Missing cancer is very harmful
Preventive Treatment 5-20% Moderate intervention risks
Major Surgery 20-50% High intervention risks
Emergency Decisions 1-15% Time-sensitive, err on caution

Analysis Type Selection

Choose Bayesian when: Choose Frequentist when:
Want probability statements Prefer traditional CI interpretation
Need EVPI calculations Simpler interpretation required
Have prior information Want to avoid prior specification
Decision uncertainty matters Bootstrap approach preferred

Sample Size Considerations

Sample Size Recommended Settings Notes
n < 100 priorStrength = 5-10 Use stronger priors
100 ≤ n < 500 priorStrength = 2-5 Moderate priors
n ≥ 500 priorStrength = 1-2 Weak priors, data-driven

Interpretation Guidelines

Understanding Net Benefit Curves

Curve Interpretation 

# Example showing different curve patterns
# (This would be actual analysis output in practice)

# Pattern 1: Clear winner
# One model consistently above others
# Interpretation: Strong evidence for using this model

# Pattern 2: Threshold-dependent
# Different models optimal at different thresholds  
# Interpretation: Threshold selection matters

# Pattern 3: Minimal differences
# All models similar performance
# Interpretation: Simple approaches may suffice

# Pattern 4: No benefit
# All models below treat all/none
# Interpretation: Models not clinically useful

Clinical Decision Rules

When to Use a Model: 1. Model curve is above both “treat all” and “treat none” 2. Net benefit difference is clinically meaningful (e.g., >0.01) 3. Confidence/credible intervals don’t include zero benefit 4. Probability of superiority >80% in relevant threshold range

When NOT to Use a Model: 1. Model performs worse than simple strategies 2. High uncertainty with wide confidence intervals 3. EVPI is very high - need more research first 4. Benefit is minimal relative to implementation costs

Statistical Significance vs Clinical Utility

Traditional Metrics vs DCA:

Traditional Approach DCA Approach
AUC = 0.75 “What’s the net benefit at my decision threshold?”
p < 0.05 “Is the improvement clinically meaningful?”
Sensitivity = 85% “Does this lead to better decisions?”
95% CI excludes 1.0 “What’s the probability this strategy is best?”

Key Point: A statistically significant model may have no clinical utility if the net benefit improvement is negligible.

Clinical Application Examples

Example 1: Cancer Screening Program 

# Evaluate biomarker panel for cancer screening
screening_analysis <- bayesdca(
  data = screening_cohort,
  outcomes = "Cancer",
  predictors = c("Biomarker_Panel", "Clinical_Score", "Imaging_Result"),
  
  # Low thresholds appropriate for screening
  thresholdMin = 0.01,     # 1% - very sensitive
  thresholdMax = 0.10,     # 10% - still screening range
  thresholdPoints = 20,
  
  bayesianAnalysis = TRUE,
  calculateEVPI = TRUE,
  
  # External prevalence from cancer registry
  useExternalPrevalence = TRUE,
  externalCases = 50,      # Registry cancer rate
  externalTotal = 1000     # Registry population
)

Key Questions: - At what threshold should we recommend further testing? - Is the biomarker panel worth the additional cost? - Should we implement this in a screening program?

Example 2: Treatment Selection Model 

# Personalized treatment selection model
treatment_analysis <- bayesdca(
  data = treatment_cohort,
  outcomes = "TreatmentFailure", 
  predictors = c("GenomicScore", "ClinicalModel", "Combined_Score"),
  
  # Moderate thresholds for treatment decisions
  thresholdMin = 0.15,     # 15% - consider treatment
  thresholdMax = 0.45,     # 45% - high-risk threshold
  thresholdPoints = 30,
  
  bayesianAnalysis = TRUE,
  priorStrength = 3,       # Moderate prior for treatment
  calculateEVPI = TRUE
)

Key Questions: - Which patients should receive intensive treatment? - Is genomic testing worth the cost for treatment selection? - What threshold should guide treatment decisions?

Example 3: Diagnostic Test Evaluation 

# Compare diagnostic tests for rapid diagnosis
diagnostic_analysis <- bayesdca(
  data = diagnostic_cohort,
  outcomes = "Disease",
  predictors = c("Rapid_Test", "Standard_Test", "Combined_Tests"),
  
  # Broad threshold range for diagnostic evaluation
  thresholdMin = 0.05,
  thresholdMax = 0.50,
  thresholdPoints = 45,
  
  bayesianAnalysis = TRUE,
  calculateEVPI = TRUE,
  
  # Direction: positive test results indicate disease
  directionIndicator = ">="
)

Key Questions: - Should we replace the standard test with the rapid test? - At what probability should we initiate treatment? - Is the combined approach worth the additional complexity?

Advanced Topics

Handling Complex Prediction Models

Probability Calibration 

# For machine learning models, ensure probabilities are well-calibrated
# before DCA analysis

# Example with logistic regression calibration
library(rms)

# Calibrate predictions
calibrated_probs <- predict(calibration_model, type = "response")

# Use calibrated probabilities in DCA
bayesdca(
  data = mydata,
  outcomes = "Death",
  predictors = "calibrated_probs",
  thresholdMin = 0.05,
  thresholdMax = 0.40,
  bayesianAnalysis = TRUE
)

Multi-class Outcomes 

# For outcomes with multiple categories, create binary versions
mydata_binary <- mydata %>%
  mutate(
    Death_vs_Others = ifelse(Outcome == "Death", 1, 0),
    Progression_vs_Others = ifelse(Outcome == "Progression", 1, 0)
  )

# Analyze each binary outcome separately
bayesdca(data = mydata_binary, outcomes = "Death_vs_Others", ...)
bayesdca(data = mydata_binary, outcomes = "Progression_vs_Others", ...)

Model Validation Considerations

Cross-validation Strategy 

# For internal validation, use cross-validation
library(pROC)

# Example 5-fold cross-validation
set.seed(123)
folds <- createFolds(mydata$Death, k = 5)

cv_results <- list()
for(i in 1:5) {
  train_data <- mydata[-folds[[i]], ]
  test_data <- mydata[folds[[i]], ]
  
  # Fit model on training data
  model <- glm(Death ~ Age + MeasurementA, data = train_data, family = binomial)
  
  # Predict on test data
  test_data$predictions <- predict(model, test_data, type = "response")
  
  # Run DCA on test predictions
  cv_results[[i]] <- bayesdca(
    data = test_data,
    outcomes = "Death", 
    predictors = "predictions",
    bayesianAnalysis = TRUE
  )
}

External Validation 

# For external validation on independent dataset
external_results <- bayesdca(
  data = external_cohort,
  outcomes = "Death",
  predictors = "model_predictions",  # From original model
  
  # May need different threshold range
  thresholdMin = 0.05,
  thresholdMax = 0.35,
  
  # Account for different prevalence
  useExternalPrevalence = TRUE,
  externalCases = external_cases,
  externalTotal = external_total,
  
  bayesianAnalysis = TRUE
)

Cost-Effectiveness Integration

Incorporating Costs 

# DCA can be extended to include costs
# Net benefit becomes: NB = (TP × Benefit - FP × Harm) / Cost

# Example: Modify net benefit calculation for cost-effectiveness
# This would require custom calculation outside the standard DCA framework

cost_adjusted_analysis <- function(dca_results, intervention_cost, harm_cost) {
  # Custom function to adjust net benefit for costs
  # Implementation would depend on specific cost structure
}

Troubleshooting and Best Practices

Common Issues and Solutions

Data Preparation Issues

Problem: Outcome variable not properly coded

# Solution: Ensure binary coding
mydata$outcome_binary <- ifelse(mydata$Outcome == "Event", 1, 0)
# Or specify positive level for factors
bayesdca(..., outcomes = "Status", outcomePos = "Dead")

Problem: Prediction probabilities outside [0,1] range

# Solution: Check and constrain predictions
summary(mydata$predictions)
mydata$predictions <- pmin(pmax(mydata$predictions, 0.001), 0.999)

Problem: Missing values in predictors

# Solution: Handle missing values appropriately
# Remove cases with missing predictors
complete_data <- mydata[complete.cases(mydata[, c("outcome", "predictor1", "predictor2")]), ]

# Or impute missing values (use appropriate method)
library(mice)
imputed_data <- complete(mice(mydata))

Threshold Selection Issues

Problem: Unclear what threshold range to use

# Solution: Use clinical guidelines or literature
# For cancer: often 5-20%
# For screening: often 1-10% 
# For surgery: often 20-50%

# Start broad and narrow based on results
bayesdca(..., thresholdMin = 0.01, thresholdMax = 0.50)

Problem: All models perform similarly

# Solutions:
# 1. Check if sample size is adequate
# 2. Consider if models are actually different
# 3. Look at EVPI - high values suggest more research needed
# 4. Focus on interpretability and implementation costs

Interpretation Challenges

Problem: Negative net benefit values

# This is normal and expected!
# Net benefit can be negative - it just means the strategy
# performs worse than "treat none"
# Focus on relative differences between strategies

Problem: Very wide confidence intervals

# Solutions:
# 1. Increase sample size if possible
# 2. Use stronger priors (increase priorStrength)
# 3. Consider if more research is needed (check EVPI)
# 4. Focus on threshold ranges with narrow intervals

Performance Optimization

Large Dataset Handling 

# For large datasets (>10,000 observations)
# Consider reducing computational burden

# Reduce posterior draws for initial exploration
bayesdca(..., nDraws = 1000)  # Instead of 2000+

# Reduce threshold points for initial analysis
bayesdca(..., thresholdPoints = 20)  # Instead of 50+

# Use frequentist approach for speed
bayesdca(..., bayesianAnalysis = FALSE, bootstrapCI = TRUE)

Memory Management 

# For memory-intensive analyses
# Clean up intermediate objects
gc()

# Use data.table for large datasets
library(data.table)
setDT(mydata)

Integration with ClinicoPath Workflow

  1. Exploratory Analysis: Use other ClinicoPath modules for initial exploration
  2. Model Development: Develop prediction models using appropriate methods
  3. Performance Evaluation: Assess discrimination (ROC analysis)
  4. Clinical Utility: Use Bayesian DCA to evaluate real-world utility
  5. Implementation Planning: Use EVPI and cost considerations for decisions

Complement with Other Modules

ROC Analysis First 

# Step 1: Traditional ROC analysis
roc_analysis <- roc(data = mydata, outcomes = "Death", predictors = "Age")

# Step 2: Decision curve analysis for clinical utility
dca_analysis <- bayesdca(data = mydata, outcomes = "Death", predictors = "Age")

Decision Trees for Implementation 

# Use decision tree analysis for implementation planning
# after DCA shows clinical utility
decision_tree_analysis <- decisiongraph(...)

Cost-Effectiveness Analysis 

# Integrate DCA results with cost-effectiveness models
# for full economic evaluation

Reporting Guidelines

Essential Elements for Publication

Methods Section

  • Specify analysis type (Bayesian vs Frequentist)
  • Report threshold range and rationale
  • Describe prior specification (if Bayesian)
  • Report software version and settings

Results Section

  • Present main decision curves with confidence intervals
  • Report net benefit at clinically relevant thresholds
  • Include probability of superiority (if Bayesian)
  • Report EVPI if calculated

Interpretation Guidelines

  • Focus on clinical meaningfulness, not just statistical significance
  • Discuss threshold-dependent findings
  • Address uncertainty and need for additional research
  • Consider implementation costs and barriers

Figure Quality

Publication-Ready Plots 

# The Bayesian DCA module produces publication-ready plots
# with proper formatting:
# - Professional color schemes
# - Percentage formatting for thresholds  
# - Appropriate confidence bands
# - Clear legends and labels

# Plots can be exported at high resolution for publications

Conclusion

Bayesian Decision Curve Analysis represents a sophisticated approach to evaluating the clinical utility of prediction models and diagnostic tests. By focusing on real-world decision making rather than just statistical performance, DCA provides crucial insights for clinical implementation.

Key Takeaways

  1. Clinical Utility Matters: A statistically significant model may have no clinical benefit
  2. Thresholds Are Critical: Performance varies dramatically across decision thresholds
  3. Uncertainty Quantification: Bayesian approaches provide richer uncertainty information
  4. Research Prioritization: EVPI helps identify where additional research is most valuable
  5. Implementation Focus: DCA directly addresses “should we use this in practice?”

When DCA Changes Clinical Practice

  • Identifies optimal decision thresholds for existing interventions
  • Prevents implementation of models with no clinical utility despite good discrimination
  • Guides resource allocation by quantifying value of additional research
  • Supports evidence-based policy for screening and diagnostic programs
  • Enables personalized medicine by identifying patient subgroups who benefit most

The ClinicoPath Bayesian DCA module provides a comprehensive toolkit for this sophisticated analysis, making advanced decision-theoretic methods accessible to clinical researchers and supporting evidence-based implementation of prediction models in healthcare.

This comprehensive guide demonstrates the full capabilities of Bayesian Decision Curve Analysis in the ClinicoPath module, providing researchers with the tools and knowledge needed to evaluate clinical utility and guide evidence-based implementation of prediction models and diagnostic tests.