meddecide: Medical Decision Analysis with Oncology Data

Introduction

This vignette demonstrates medical decision analysis using the meddecide module with oncology datasets from the OncoDataSets package. We’ll cover diagnostic test evaluation, ROC analysis, decision curves, and decision tree modeling.

Loading Required Packages


library(OncoDataSets)
library(dplyr)
library(knitr)

Example 1: PSA for Prostate Cancer Detection

The PSA dataset contains prostate-specific antigen measurements and cancer outcomes.

data("PSAProstateCancer_df")

# Dataset overview
str(PSAProstateCancer_df)
#> 'data.frame':    97 obs. of  9 variables:
#>  $ lcavol : num  -0.58 -0.994 -0.511 -1.204 0.751 ...
#>  $ lweight: num  2.77 3.32 2.69 3.28 3.43 ...
#>  $ age    : int  50 58 74 58 62 50 64 58 47 63 ...
#>  $ lbph   : num  -1.39 -1.39 -1.39 -1.39 -1.39 ...
#>  $ svi    : int  0 0 0 0 0 0 0 0 0 0 ...
#>  $ lcp    : num  -1.39 -1.39 -1.39 -1.39 -1.39 ...
#>  $ gleason: int  6 6 7 6 6 6 6 6 6 6 ...
#>  $ pgg45  : int  0 0 20 0 0 0 0 0 0 0 ...
#>  $ lpsa   : num  -0.431 -0.163 -0.163 -0.163 0.372 ...

# Create binary outcome for high-grade cancer
PSAProstateCancer_df$high_grade <- ifelse(PSAProstateCancer_df$gleason >= 7, 1, 0)

# PSA levels (back-transform from log)
PSAProstateCancer_df$psa <- exp(PSAProstateCancer_df$lpsa)

summary(PSAProstateCancer_df$psa)
#>    Min. 1st Qu.  Median    Mean 3rd Qu.    Max. 
#>    0.65    5.65   13.35   23.74   21.25  265.85
table(PSAProstateCancer_df$high_grade)
#> 
#>  0  1 
#> 35 62

ROC Analysis for PSA

# In jamovi:
# 1. Select Analyses → meddecide → ROC Analysis
# 2. Set 'psa' as test variable
# 3. Set 'high_grade' as state variable (1 = disease)
# 4. Enable:
#    - ROC curve with confidence bands
#    - Optimal cutoff calculation
#    - Sensitivity/specificity table
#    - Likelihood ratios

Multiple Biomarker Comparison

Compare PSA with other clinical factors:

# In jamovi:
# 1. Select Analyses → meddecide → ROC Comparison
# 2. Add multiple test variables:
#    - psa
#    - age
#    - lweight (log prostate weight)
# 3. Compare AUCs with DeLong test
# 4. Generate comparative ROC plot

Decision Curve Analysis

Evaluate clinical utility across different threshold probabilities:

# In jamovi:
# 1. Select Analyses → meddecide → Decision Curve
# 2. Add PSA-based model
# 3. Compare with:
#    - Treat all strategy
#    - Treat none strategy
# 4. Set threshold probability range (0-50%)
# 5. Calculate net benefit

Example 2: CA19-9 for Pancreatic Cancer

Meta-analysis data for CA19-9 diagnostic accuracy:

data("CA19PancreaticCancer_df")

# Dataset structure (diagnostic accuracy studies)
str(CA19PancreaticCancer_df)
#> Classes 'data.table' and 'data.frame':   22 obs. of  5 variables:
#>  $ study: chr  "Andriulli" "Benini" "DelFavero" "Gupta" ...
#>  $ TP   : int  64 23 18 13 74 27 20 11 19 32 ...
#>  $ FP   : int  148 7 20 64 23 27 39 5 49 10 ...
#>  $ FN   : int  12 2 6 4 21 13 4 3 6 5 ...
#>  $ TN   : int  294 118 29 589 83 265 103 43 172 29 ...

# Calculate sensitivity and specificity for each study
CA19PancreaticCancer_df <- CA19PancreaticCancer_df %>%
  mutate(
    sensitivity = TP / (TP + FN),
    specificity = TN / (TN + FP),
    total_n = TP + FP + FN + TN
  )

kable(CA19PancreaticCancer_df[1:5, c("study", "sensitivity", "specificity", "total_n")])

study	sensitivity	specificity	total_n
Andriulli	0.8421053	0.6651584	518
Benini	0.9200000	0.9440000	150
DelFavero	0.7500000	0.5918367	73
Gupta	0.7647059	0.9019908	670
Haglund	0.7789474	0.7830189	201

Meta-analysis of Diagnostic Accuracy

# In jamovi:
# 1. Select Analyses → meddecide → Diagnostic Meta-analysis
# 2. Input TP, FP, FN, TN for each study
# 3. Enable:
#    - Summary ROC curve
#    - Heterogeneity assessment
#    - Forest plots
#    - Publication bias analysis

Bayesian Analysis

Calculate post-test probabilities:

# In jamovi:
# 1. Select Analyses → meddecide → Bayesian Calculator
# 2. Input pooled sensitivity/specificity
# 3. Set pre-test probabilities (prevalence)
# 4. Calculate:
#    - Positive predictive value
#    - Negative predictive value
#    - Post-test probability curves

Example 3: Lung Nodule Evaluation

data("LungNodulesDetected_df")

# Dataset overview
str(LungNodulesDetected_df)
#> Classes 'data.table' and 'data.frame':   999 obs. of  8 variables:
#>  $ sex          : Factor w/ 2 levels "F","M": 2 2 1 1 2 2 2 1 1 2 ...
#>  $ age          : num  84 53.9 75.5 66.1 47.8 38.6 77.9 77.5 46.6 67.8 ...
#>  $ num.annotated: num  0 0 0 0 0 0 0 0 0 0 ...
#>  $ location     : Factor w/ 6 levels "Lingular Segment",..: 5 1 5 5 4 1 2 3 3 6 ...
#>  $ spiculate    : Factor w/ 2 levels "No","Yes": 1 1 1 1 1 1 1 2 1 1 ...
#>  $ smoke.status : Factor w/ 5 levels "current","exsmoke",..: 4 4 4 4 4 4 4 4 4 4 ...
#>  $ diameter     : num  5 7 8 5 5 6 7 11 5 10 ...
#>  $ malignant    : num  0 0 0 0 0 0 0 0 0 0 ...
#>  - attr(*, ".internal.selfref")=<externalptr>

# Malignancy by nodule characteristics
table(LungNodulesDetected_df$spiculate, LungNodulesDetected_df$malignant)
#>      
#>         0   1
#>   No  665  85
#>   Yes 110 139
table(cut(LungNodulesDetected_df$diameter, c(0, 5, 10, 20, Inf)), 
      LungNodulesDetected_df$malignant)
#>           
#>              0   1
#>   (0,5]    166   7
#>   (5,10]   506  70
#>   (10,20]  103 147
#>   (20,Inf]   0   0

Decision Tree for Nodule Management

# In jamovi:
# 1. Select Analyses → meddecide → Decision Tree
# 2. Create decision nodes:
#    - Nodule detected
#    - Size > 8mm?
#    - Spiculated?
#    - Biopsy vs. follow-up
# 3. Input probabilities:
#    - Malignancy rates by size/features
#    - Test characteristics (biopsy sensitivity/specificity)
# 4. Input utilities/costs:
#    - Early detection benefit
#    - Biopsy complications
#    - Anxiety from follow-up
# 5. Calculate expected values

Clinical Decision Rule Development

# In jamovi:
# 1. Use Classification Analysis
# 2. Predictors: diameter, spiculate, age, smoking
# 3. Outcome: malignant
# 4. Methods:
#    - Logistic regression
#    - Decision tree (CART)
#    - Random forest
# 5. Validate with cross-validation
# 6. Create clinical scoring system

Example 4: Comparative Effectiveness

Using multiple datasets to compare diagnostic strategies:

Prostate Cancer Screening Strategies

# In jamovi:
# 1. Select Analyses → meddecide → Decision Compare
# 2. Define strategies:
#    - PSA alone (cutoff 4.0)
#    - PSA + age adjustment
#    - PSA + free/total ratio
#    - MRI-based screening
# 3. Input:
#    - Test characteristics
#    - Costs
#    - Quality-adjusted life years
# 4. Perform cost-effectiveness analysis
# 5. Generate efficiency frontier

Example 5: Markov Model for Cancer Progression

Long-term outcomes modeling:

# In jamovi:
# 1. Select Analyses → meddecide → Decision Tree (Markov mode)
# 2. Define health states:
#    - No cancer
#    - Localized cancer
#    - Advanced cancer
#    - Death
# 3. Input transition probabilities
# 4. Set cycle length (e.g., 1 year)
# 5. Time horizon (e.g., lifetime)
# 6. Calculate:
#    - Life expectancy
#    - Quality-adjusted life years
#    - Incremental cost-effectiveness

Practical Applications

1. Biomarker Cutoff Optimization

# Workflow:
# 1. ROC analysis to assess discrimination
# 2. Calculate optimal cutoffs for:
#    - Youden index (sensitivity + specificity)
#    - Clinical utility (weighted by consequences)
#    - Cost-effectiveness
# 3. Validate in independent dataset

2. Risk Prediction Model

# Using PSA data:
# 1. Develop multivariable model
# 2. Generate risk scores
# 3. Create risk categories
# 4. Decision curve analysis
# 5. Clinical impact assessment:
#    - Number needed to screen
#    - False positives avoided
#    - Cancers detected

3. Diagnostic Pathway Design

# Example: Lung nodule workup
# 1. Initial CT findings
# 2. Risk stratification (size, features)
# 3. Decision points:
#    - Immediate biopsy
#    - PET scan
#    - Short-term follow-up
#    - Annual screening
# 4. Calculate pathway efficiency

Advanced Features

Sensitivity Analysis

# In jamovi meddecide:
# 1. One-way sensitivity analysis
#    - Vary each parameter
#    - Identify threshold values
# 2. Two-way sensitivity analysis
#    - Vary two parameters simultaneously
#    - Create contour plots
# 3. Probabilistic sensitivity analysis
#    - Monte Carlo simulation
#    - Confidence ellipses

Time-to-Event Integration

# Combine with survival data:
# 1. Time-dependent ROC curves
# 2. Dynamic decision curves
# 3. Landmark analysis
# 4. Risk prediction over time

Best Practices

Define Clinical Question: Clear objectives for decision analysis
Evidence Quality: Document sources for probabilities/utilities
Perspective: Specify (patient, provider, societal)
Uncertainty: Always perform sensitivity analyses
Validation: Test models in independent data

Reporting Guidelines

For Diagnostic Studies

STARD checklist compliance
Report all performance metrics
Confidence intervals
Clinical utility measures

For Decision Analysis

Model structure diagram
Parameter sources
Base case results
Sensitivity analyses
Clinical implications

Integration Example

Complete diagnostic evaluation workflow:

# 1. Data Preparation
#    - Load OncoDataSets
#    - Define outcomes
#    - Handle missing data

# 2. Test Performance
#    - ROC analysis
#    - Optimal cutoffs
#    - Subgroup analyses

# 3. Clinical Utility
#    - Decision curves
#    - Net benefit calculation
#    - Number needed to test

# 4. Decision Modeling
#    - Build decision tree
#    - Input probabilities
#    - Calculate expected values

# 5. Sensitivity Analysis
#    - Test robustness
#    - Identify key drivers

# 6. Implementation
#    - Clinical algorithm
#    - Decision support tool
#    - Monitoring plan

Conclusion

The meddecide module provides comprehensive tools for medical decision analysis in oncology:

Evaluate diagnostic test performance
Compare multiple biomarkers
Assess clinical utility beyond accuracy
Model complex decision scenarios
Support evidence-based clinical guidelines

Combined with OncoDataSets, researchers can develop and validate decision support tools using real-world cancer data.

References

OncoDataSets package: https://cran.r-project.org/package=OncoDataSets
meddecide documentation: https://www.serdarbalci.com/meddecide/
Decision curve analysis: https://www.mskcc.org/departments/epidemiology-biostatistics/biostatistics/decision-curve-analysis

ClinicoPath Team

2025-10-09