Machine Learning versus Bayesian Spatial Models for Small-Area Tuberculosis Prediction: A Comparative Study in Metropolitan Santiago, Chile

Draft in preparation

Reproducible R pipeline for district-level TB rate prediction in Metropolitan Santiago, Chile, comparing negative binomial spatial lag, tree-based ML (RF-S, BART, XGBoost), BYM2 (R-INLA), and NNLS stacking.

Main figures

Figure 1. Study area and spatial distribution of tuberculosis outcomes in Metropolitan Santiago

Location of the study area in Chile (panel A), district-level TB case counts (panel B), and TB rate per 10,000 inhabitants (panel C).

Figure 2. Out-of-fold predicted TB rates by model

Choropleth maps and observed-vs-predicted scatter plots for NB-lag, BART, RF-S, XGB, and BYM2 (pooled OOF evaluation).

Figure 3. Cross-model predictive performance (pooled out-of-fold)

Comparison of MAE, RMSE, BIAS, R², CE, and IA across NB-lag, BART, RF-S, XGB, and BYM2.

Figure 4. Stacking ensemble: spatial predictions, residuals, and model comparison

OOF predicted rates and residuals for the NNLS stacking meta-learner, with fit metrics compared against XGB and RF-S.

Funding

FONIS Nº SA24I0203

Authors

Role	Name	Affiliation	Contact
Co-author / repository manager	José Daniel Conejeros	ISI Foundation, Turin; SENTINET	jdconejeros@uc.cl
Corresponding author	Salvador Ayala	Centro de Epidemiología y Políticas de Salud, Facultad de Medicina Clínica Alemana, Universidad del Desarrollo, Santiago	salvadorayala@udd.cl

Project overview

Background

Tuberculosis (TB) burden is spatially heterogeneous. Classical disease-mapping approaches (Poisson/NB regression, CAR/BYM models) explicitly model spatial structure but may underfit complex socio-environmental gradients. Machine learning (ML) algorithms capture non-linearities but often treat space indirectly. This repository implements a reproducible comparative pipeline for small-area TB rate prediction in the Metropolitan Region of Santiago (RM), Chile, combining:

• Frequentist spatial regression (NB with spatial lag)
• Tree-based ML (RF-S, BART, XGBoost)
• Bayesian hierarchical spatial models (BYM2 via R-INLA)
• Ensemble stacking (NNLS meta-learner on OOF base-model predictions)

Unit of analysis

• 343 urban districts with polygon geometry
• Outcome: casos_tb; rate metric tasa_tb_10k (per 10,000 population)
• Queen contiguity spatial weights; train-only spatial lags in cross-validation

Objectives

1. Build a harmonized district-level analytic dataset (census, health access, climate).
2. Fit and cross-validate multiple spatial and ML models on a common covariate pool.
3. Evaluate predictive performance (MAE, RMSE, BIAS, CE, IA, R²) and residual spatial structure (Moran's I, LISA).
4. Produce publication-ready maps, diagnostic panels, variable-importance figures, and tables.

Full statistical and model specifications: 02_code/03_models/README.md.

---

Repository structure

TB_ML_FONIS/
├── 01_data/
│   ├── raw/              Census, health facilities, climate extracts
│   ├── interim/          Queen spatial weights (nb, lw)
│   └── analysis/         District analytic table (sf + covariates)
├── 02_code/
│   ├── 00_setup/         Paths, map utilities, dependencies
│   ├── 01_prepare/       Dataset construction (01a–01d)
│   ├── 02_descriptives/  Exploratory maps and correlation panels
│   ├── 03_models/        Model fitting (04a–04j) + figures/tables (05)
│   └── python/           Earth Engine climate extraction
├── 03_output/
│   ├── maps/             Descriptive choropleths
│   ├── figures/          Descriptive panels
│   └── models/           Model RDS, metrics CSV, Figures/, Tables/
├── 04_Paper/             Manuscript and supplementary files
└── README.md             This file

---

Reproducibility

System requirements

• R ≥ 4.3 (4.5+ recommended)
• Optional: Python 3 + Earth Engine API for climate re-extraction

Install R dependencies:

Rscript 02_code/00_setup/install_spatial_ml_dependencies.R

INLA and fastshap require additional setup; see install_spatial_ml_dependencies.R.

Pipeline execution

1. Build analytic dataset

Rscript 02_code/01_prepare/01a_prepare_census_covariates.R
Rscript 02_code/01_prepare/01b_prepare_health_covariates.R
Rscript 02_code/01_prepare/01c_prepare_climate_covariates.R
Rscript 02_code/01_prepare/01d_prepare_analysis_dataset.R

2. Descriptive analysis

Rscript 02_code/02_descriptives/02_descriptive_analysis.R

3. Spatial models (run order in 02_code/03_models/README.md)

Rscript 02_code/03_models/04a_variable_selection_regularization.R
Rscript 02_code/03_models/04b_nb_lag_model.R
Rscript 02_code/03_models/04c_rf_s_model.R
Rscript 02_code/03_models/04d_bart_car_model.R
Rscript 02_code/03_models/04e_xgb_model.R
Rscript 02_code/03_models/04f_inla_model.R
Rscript 02_code/03_models/04g_metamodel_stacking.R
Rscript 02_code/03_models/04h_model_parametrization_table.R
Rscript 02_code/03_models/04j_spatial_residual_diagnostics.R
Rscript 02_code/03_models/05_visualize_spatial_models.R

4. Climate extraction (optional refresh)

python 02_code/python/06_temperature_ndvi_extraction.py --dry-run
python 02_code/python/06_temperature_ndvi_extraction.py

Approximate total runtime

Stage	Time
Data preparation (01)	< 5 min
Descriptives (02)	< 10 min
Models 04a–04d, 04g–04j	< 30 min
XGB (04e)	10–30 min
BYM2 (04f)	5–15 min
Publication figures (05)	10–20 min

---

Key outputs

Location	Content
01_data/analysis/tb_analisis_distritos_rm.rds	Analysis dataset
03_output/models/model_*.rds	Fitted models + OOF predictions
03_output/models/metrics_all_models.csv	Cross-model metrics
03_output/models/Figures/	Publication figure panels
03_output/models/Tables/	Publication tables (CSV + XLSX)
03_output/maps/	Descriptive maps

Main figure panels (05)

Panel	File (example)
Predicted rates (OOF)	Figures/predicted/panel_predicted_rates_oof_pooled.png
Residual maps	Figures/residual_diagnostics/panel_residual_maps_oof_pooled.png
LISA clusters	Figures/residual_diagnostics/panel_lisa_maps_oof_pooled.png
Fit metrics	Figures/fit/panel_fit_metrics_oof_pooled.png
SHAP + importance	Figures/variables_analysis/panel_shap_importance.png
Stacking summary	Figures/stacking/panel_stacking_oof_pooled.png

---

Variables

Outcomes

Variable	Description	Source
casos_tb	Notified TB cases by district	TB surveillance system
tasa_tb_10k	Rate per 10,000 inhabitants	Derived from cases and population

Covariates (selection)

Domain	Variables (examples)	Source
Demographics	Population density, sex/age structure, migration, indigenous peoples, overcrowding, mean age	Census 2024 (district aggregates)
Health access	APS establishment density, distance to nearest primary-care facility	IDE Chile / Geoportal (Establecimientos de salud de Chile, Febrero 2026)
Climate	Mean NDVI, mean/max/min temperature (2024)	Google Earth Engine

Spatial lag of TB rate (train-only in CV) is added as a predictor for NB-lag, RF-S, BART, and XGB.

---

Model documentation

All implemented models, mathematical specifications, cross-validation design, metrics, and output contracts are documented in:

→ 02_code/03_models/README.md

Models in the main comparison: NB-lag, BART, RF-S, XGB, BYM2, Stacking. The INLA VIF-reduced variant is supplementary.

---

Data availability

Individual-level TB notification data are subject to Chilean health confidentiality rules. District-level aggregates used in this repository are derived from administrative sources described above. Processed analytic files are stored under 01_data/analysis/ when available in the workspace.

---

Acknowledgments

Supported by FONIS Nº SA24I0203. We thank the Chilean Ministry of Health and census data providers. Climate covariates use Google Earth Engine extracts (see 02_code/python/).

---

Contact

• Salvador Ayala (corresponding): salvadorayala@udd.cl
• José Daniel Conejeros (code): jdconejeros@uc.cl