BMC2025: Risk prediction models for depression in older adults with cancer

Overview

TL;DR

• This repository contains an R-based predictive modeling workflow for long-term depression risk estimation (euro_d) from longitudinal clinical data.
• Project framing targets a 2-year prediction horizon in a clinical follow-up setting.
• Models compared in code: Decision Tree (rpart), Random Forest (rf), and Generalized Linear Model (glm).
• Feature selection strategies implemented: random forward selection, random backward selection, and genetic algorithms.
• Validation includes k-fold cross-validation and repeated double cross-validation (double_cv in cv.R), with threshold analysis.
• Data artifacts in this repo include high-dimensional longitudinal tables up to 42669 x 283 (data/not_onco_combined_dataset.sav), while the primary modeling dataset used in the notebook pipeline is 4057 x 89 (data/dataset.sav).
• Key screening-oriented metrics (reported in the associated study context): AUC 0.81, Sensitivity 90.7%, Specificity 43.8%.

Model Performance

ROC curve comparing model discrimination performance (GA-selected feature setting).

Calibration curve showing agreement between predicted and observed outcomes.

Precision-Recall curve highlighting performance under class imbalance.

Methodology

• Supervised learning task: binary depression-risk prediction on target variable euro_d.
• Longitudinal clinical data: wave-based variables processed and harmonized across follow-up.
• Feature selection: random forward selection (fwfs.R), random backward selection (bwfs.R), and genetic feature selection (genetic.R).
• Validation design: k-fold cross-validation plus repeated double cross-validation (double_cv, configurable under/over-sampling).
• Model comparison: Decision Tree (rpart), Random Forest (rf), and GLM (glm) with threshold-dependent operating-point analysis.

Clinical Use Case

The intended use is early risk screening for depression in clinical populations (including oncology-focused cohorts in this codebase).
Given a screening objective, sensitivity is prioritized to reduce false negatives and support earlier follow-up evaluation, while accepting lower specificity as a triage tradeoff.

Repository Structure

• data/: input SPSS datasets (dataset.sav, score_dataset.sav, not_onco_combined_dataset.sav).
• scripts/: setup and environment validation (setup_environment.R, check_environment.R).
• notebooks/: exploratory and reporting workflows (analisi_base.ipynb, analisi_base.qmd).
• results/: exported model-performance figures (ROC, calibration, PR).
• predictions/: saved out-of-fold prediction outputs across balancing/selection strategies.
• thresholds/: threshold-level sensitivity/specificity/PPV/NPV summaries.
• Root R scripts (pipeline.R, cv.R, fwfs.R, bwfs.R, genetic.R, ddop_*.R): core modeling and utility pipeline.

How to Run

Minimal local workflow from repository root:

Rscript scripts/setup_environment.R --register-kernel
Rscript scripts/check_environment.R
python -m jupyter lab
quarto render notebooks/analisi_base.qmd

Then open:

• notebooks/analisi_base.ipynb (R kernel), or
• notebooks/analisi_base.html after Quarto rendering.

To execute the full pipeline programmatically, use pipeline.R (entry function rpm()).

If command details need adjustment for your environment, use repository entry points (scripts/, notebooks/, and pipeline.R) as the source of truth rather than generic external commands.

Online options already used in this project context:

• Posit Cloud: upload the project folder and render notebooks/analisi_base.qmd.
• GitHub Codespaces: clone the repository and run the same commands.

Reference

Belvederi Murri M, Sciavicco G, Specchia M, et al. Risk prediction models for depression in older adults with cancer. BMC Psychiatry. 2025.