Introduction to Decision Panel Optimization

meddecide Development Team

2025-07-29

Introduction

The Decision Panel Optimization module in the meddecide package provides a comprehensive framework for optimizing diagnostic test combinations in medical decision-making. This vignette introduces the basic concepts and demonstrates core functionality.

Key Concepts

Testing Strategies

When multiple diagnostic tests are available, they can be combined in different ways:

1. Single Testing: Use individual tests independently
2. Parallel Testing: Perform multiple tests simultaneously
   • ANY rule (OR): Positive if any test is positive
   • ALL rule (AND): Positive only if all tests are positive
   • MAJORITY rule: Positive if majority of tests are positive
3. Sequential Testing: Perform tests in sequence based on previous results
   • Stop on first positive
   • Confirmatory (require multiple positives)
   • Exclusion (require multiple negatives)

Optimization Criteria

The module can optimize test panels based on various criteria:

• Accuracy: Overall correct classification rate
• Sensitivity: Ability to detect disease (minimize false negatives)
• Specificity: Ability to rule out disease (minimize false positives)
• Predictive Values: PPV and NPV
• Cost-Effectiveness: Balance performance with resource utilization
• Utility: Custom utility functions incorporating costs of errors

Installation and Loading

# Install meddecide package
install.packages("meddecide")

# Or install from GitHub
devtools::install_github("ClinicoPath/meddecide")

# Load required packages
library(meddecide)
library(dplyr)
library(ggplot2)
library(rpart)
library(rpart.plot)
library(knitr)
library(forcats)

Basic Example: COVID-19 Screening

Let’s start with a simple example using COVID-19 screening data:

# Examine the data structure
str(covid_screening_data)

# Check disease prevalence
table(covid_screening_data$covid_status)
prop.table(table(covid_screening_data$covid_status))

Running Basic Analysis

# Basic decision panel analysis
covid_panel <- decisionpanel(
  data = covid_screening_data,
  tests = c("rapid_antigen", "pcr", "chest_ct"),
  testLevels = c("Positive", "Positive", "Abnormal"),
  gold = "covid_status",
  goldPositive = "Positive",
  strategies = "all",
  optimizationCriteria = "accuracy"
)

Interpreting Results

The analysis provides several key outputs:

1. Individual Test Performance: How each test performs alone
2. Optimal Panel: The best combination of tests
3. Strategy Comparison: Performance of different testing approaches
4. Decision Tree: Optimal sequence for testing

Understanding Testing Strategies

Parallel Testing Example

# Simulate parallel testing with ANY rule
# Positive if rapid_antigen OR pcr is positive
parallel_any <- with(covid_screening_data, 
  rapid_antigen == "Positive" | pcr == "Positive"
)

# Create confusion matrix
conf_matrix_any <- table(
  Predicted = parallel_any,
  Actual = covid_screening_data$covid_status == "Positive"
)

print(conf_matrix_any)

# Calculate metrics
sensitivity_any <- conf_matrix_any[2,2] / sum(conf_matrix_any[,2])
specificity_any <- conf_matrix_any[1,1] / sum(conf_matrix_any[,1])

cat("Parallel ANY Rule:\n")
cat(sprintf("Sensitivity: %.1f%%\n", sensitivity_any * 100))
cat(sprintf("Specificity: %.1f%%\n", specificity_any * 100))

Sequential Testing Example

# Simulate sequential testing
# Start with rapid test, only do PCR if rapid is positive
sequential_result <- rep("Negative", nrow(covid_screening_data))

# Those with positive rapid test
rapid_pos_idx <- which(covid_screening_data$rapid_antigen == "Positive")

# Among those, check PCR
sequential_result[rapid_pos_idx] <- 
  ifelse(covid_screening_data$pcr[rapid_pos_idx] == "Positive", 
         "Positive", "Negative")

# Create confusion matrix
conf_matrix_seq <- table(
  Predicted = sequential_result == "Positive",
  Actual = covid_screening_data$covid_status == "Positive"
)

print(conf_matrix_seq)

# Calculate metrics
sensitivity_seq <- conf_matrix_seq[2,2] / sum(conf_matrix_seq[,2])
specificity_seq <- conf_matrix_seq[1,1] / sum(conf_matrix_seq[,1])

cat("\nSequential Testing:\n")
cat(sprintf("Sensitivity: %.1f%%\n", sensitivity_seq * 100))
cat(sprintf("Specificity: %.1f%%\n", specificity_seq * 100))

# Calculate cost savings
pcr_tests_saved <- sum(covid_screening_data$rapid_antigen == "Negative")
cat(sprintf("PCR tests saved: %d (%.1f%%)\n", 
            pcr_tests_saved, 
            pcr_tests_saved/nrow(covid_screening_data) * 100))

Cost-Effectiveness Analysis

When costs are considered, the optimal strategy may change:

# Analysis with costs
covid_panel_cost <- decisionpanel(
  data = covid_screening_data,
  tests = c("rapid_antigen", "pcr", "chest_ct"),
  testLevels = c("Positive", "Positive", "Abnormal"),
  gold = "covid_status",
  goldPositive = "Positive",
  strategies = "all",
  optimizationCriteria = "utility",
  useCosts = TRUE,
  testCosts = "5,50,200",  # Costs for each test
  fpCost = 500,            # Cost of false positive
  fnCost = 5000            # Cost of false negative
)

Visualization

Performance Comparison Plot

# Create performance comparison data
strategies <- data.frame(
  Strategy = c("Rapid Only", "PCR Only", "Parallel ANY", "Sequential"),
  Sensitivity = c(65, 95, 98, 62),
  Specificity = c(98, 99, 97, 99.9),
  Cost = c(5, 50, 55, 15)
)

# Plot sensitivity vs specificity
ggplot(strategies, aes(x = 100 - Specificity, y = Sensitivity)) +
  geom_point(aes(size = Cost), alpha = 0.6) +
  geom_text(aes(label = Strategy), vjust = -1) +
  scale_size_continuous(range = c(3, 10)) +
  xlim(0, 5) + ylim(60, 100) +
  labs(
    title = "Testing Strategy Comparison",
    x = "False Positive Rate (%)",
    y = "Sensitivity (%)",
    size = "Cost ($)"
  ) +
  theme_minimal()

Decision Trees

Decision trees provide clear algorithms for clinical use:

# Generate decision tree
covid_tree <- decisionpanel(
  data = covid_screening_data,
  tests = c("rapid_antigen", "pcr", "chest_ct", "symptom_score"),
  testLevels = c("Positive", "Positive", "Abnormal", ">5"),
  gold = "covid_status",
  goldPositive = "Positive",
  createTree = TRUE,
  treeMethod = "cart",
  maxDepth = 3
)

Interpreting the Tree

A typical decision tree output might look like:

1. Start with Rapid Antigen Test
   ├─ If Positive (2% of patients)
   │  └─ Confirm with PCR
   │     ├─ If Positive → COVID Positive (PPV: 95%)
   │     └─ If Negative → COVID Negative (NPV: 98%)
   └─ If Negative (98% of patients)
      ├─ If Symptoms > 5
      │  └─ Perform Chest CT
      │     ├─ If Abnormal → Perform PCR
      │     └─ If Normal → COVID Negative
      └─ If Symptoms ≤ 5 → COVID Negative

Advanced Features

Cross-Validation

Validate panel performance using k-fold cross-validation:

# Run with cross-validation
covid_panel_cv <- decisionpanel(
  data = covid_screening_data,
  tests = c("rapid_antigen", "pcr", "chest_ct"),
  testLevels = c("Positive", "Positive", "Abnormal"),
  gold = "covid_status",
  goldPositive = "Positive",
  crossValidate = TRUE,
  nFolds = 5,
  seed = 123
)

Bootstrap Confidence Intervals

Get uncertainty estimates for performance metrics:

# Run with bootstrap
covid_panel_boot <- decisionpanel(
  data = covid_screening_data,
  tests = c("rapid_antigen", "pcr", "chest_ct"),
  testLevels = c("Positive", "Positive", "Abnormal"),
  gold = "covid_status",
  goldPositive = "Positive",
  bootstrap = TRUE,
  bootReps = 1000,
  seed = 123
)

Best Practices

1. Start Simple: Begin with individual test performance before combinations
2. Consider Context: Screening vs. diagnosis requires different strategies
3. Validate Results: Use cross-validation or separate test sets
4. Include Costs: Real-world decisions must consider resources
5. Think Sequentially: Often more efficient than parallel testing
6. Set Constraints: Define minimum acceptable performance
7. Interpret Clinically: Statistical optimality isn’t everything

Conclusion

The Decision Panel Optimization module provides a systematic approach to combining diagnostic tests. By considering various strategies, costs, and constraints, it helps identify practical testing algorithms that balance performance with resource utilization.

Next Steps

• See the “Clinical Applications” vignette for disease-specific examples
• Review “Advanced Optimization” for complex scenarios
• Check “Implementation Guide” for deploying algorithms in practice

Session Information

sessionInfo()