A portfolio case study in validation-first analytics using ClinicalTrials.gov oncology trials

This project demonstrates end-to-end predictive analytics for clinical trial duration modeling—and why validation strategy matters more than model selection.

Initial random validation suggested strong performance (R² = 0.84). A time-based validation split—appropriate for prospective use—revealed minimal predictive power (R² ≈ 0.04).

The finding is the insight: realistic validation prevents overconfident forecasting from registry data.

• Temporal drift: Trial durations shortened substantially across eras (median ~82 → ~53 months in this dataset)
• Registry limitations: ClinicalTrials.gov lacks protocol complexity, endpoint strategy, and competitive dynamics
• Evolving practices: Adaptive designs, biomarker selection, and regulatory changes altered duration drivers over time

This is not "model failure"—it shows why validation must match the intended decision context.

Key metrics with honest R² = 0.04 and model summary

Interactive tool showing how the model behaves—not for operational planning

Coefficient analysis and feature distributions

Diagnostic plots with actual error analysis by phase and sponsor

Validation Lessons panel explaining why metrics can mislead

The dashboard helps answer:

• How do registry-level trial characteristics relate to historical duration?
• What are the limits of registry-based forecasting in oncology?
• How does validation choice change conclusions?
• Where does the model fail, and why?

This tool is intended for methodology and pattern understanding—not operational planning.

To avoid overreach, this project does not:

• Provide reliable prospective duration forecasts for planning
• Recommend protocol or design decisions
• Predict clinical success, approvals, or outcomes
• Substitute for protocol-level feasibility analysis

All results are conditional on historical registry patterns that may not persist.

1. Acquisition: ClinicalTrials.gov API v2 (pulled January 2026)
2. Cohort: Interventional, treatment-purpose oncology trials (keyword-based condition query; imperfect)
3. Outcome: Duration in months from Start Date → Completion Date (includes follow-up; not equivalent to LPLV)
4. Features: Enrollment, sites, countries, sponsor type, phase, derived ratios/log transforms
5. Model: Elastic Net (glmnet) via tidymodels

Training:  Trials with start date ≤2015 (n=546)
Test:      Trials with start date 2016–2018 (n=340)
Excluded:  Trials with start date ≥2019 (n=878, right-truncation bias)

Trials starting in 2019+ were excluded because only shorter-duration studies from that era are fully completed and present in the dataset as of the January 2026 data pull. Including them would bias observed durations downward.

Final model used in the dashboard: Elastic Net (glmnet)

• ClinicalTrials.gov (U.S. National Library of Medicine)
• API v2, accessed January 2026

• Interventional studies
• Primary purpose: treatment
• Phase 1/2, 2, 2/3, or 3
• Completed with valid start/completion dates
• Actual enrollment (final posted value)
• Duration 1–180 months

• Registry fields are self-reported and updated over time
• Historical snapshots are not available in this build
• Oncology cohort defined via keyword-based registry fields (imperfect)

• Validation-first thinking — Match strategy to use case
• Transparency over performance — Report honest metrics
• Explainability over complexity — Linear model over black box
• Appropriate restraint — Document what the tool cannot do
• Decision support, not automation — Inform, don't prescribe

• Oncology cohort defined via keyword-based registry fields (imperfect)
• Registry fields are self-reported and updated over time
• Limited prospective predictive power (R² ≈ 0.04)
• Temporal drift may continue beyond observed patterns
• Survivorship bias (completed trials only)

clinical-trial-forecaster/
├── app/
│   ├── app.R                    # Entry point
│   ├── global.R                 # Data loading, metrics, helpers
│   ├── ui.R                     # Main UI definition (moved from R/)
│   ├── DESCRIPTION              # Package dependencies for shinyapps.io
│   ├── modules/
│   │   ├── mod_executive_brief.R
│   │   ├── mod_forecaster.R
│   │   ├── mod_features.R
│   │   ├── mod_performance.R
│   │   └── mod_methods.R
│   ├── models/                  # Trained model artifacts
│   │   ├── final_workflow.rds
│   │   ├── feature_importance.rds
│   │   ├── model_comparison.csv
│   │   └── test_predictions.csv
│   ├── data-raw/                # Trial data
│   └── www/
│       └── styles.css
├── scripts/                     # Data pipeline
│   ├── 01_data_acquisition.R
│   ├── 02_data_exploration.R
│   ├── 03_feature_engineering.R
│   └── 04_modeling.R
├── docs/                        # Diagnostic plots
└── screenshots/

• Language: R
• Framework: Shiny (modular architecture)
• UI Components: shiny.semantic (Appsilon)
• Visualization: ggplot2, ggiraph, reactable
• ML Framework: tidymodels
• Deployment: ShinyApps.io

This project is released under the MIT License.

This project is for educational and portfolio purposes only.

It does not constitute clinical, operational, or planning advice. All parameters are derived from publicly available registry data. No representation is made regarding applicability to specific trial contexts.

Steven Ponce Data Analyst | R Shiny Developer | Business Intelligence Specialist

🔗 Portfolio Website: stevenponce.netlify.app 🐙 GitHub: @poncest 💼 LinkedIn: stevenponce 🦋 Bluesky: @sponce1

Prepared by Steven Ponce as part of a professional analytics portfolio. Demonstrating that honest validation findings—even when they reveal limitations—are more valuable than inflated metrics.