⚠️ Educational project — not a clinical tool, not medical advice. National, aggregate estimates only.
WHO 2021 · IHME GBD 2023 · World Bank

Turning global mental-health data into clear, honest insight.

A full-stack analytics project: reproducible pipelines, leakage-free machine learning, data mining, deep-learning forecasting, and a live predictor of national suicide rates — built for transparency, not hype.

Try the predictorExplore the dashboardsCode on GitHub
scikit-learnLightGBM · OptunaSHAPconformal intervalsUMAP · networkxN-BEATSFastAPI · DockerNext.js

What the data actually says

Findings from the real WHO + IHME + World Bank data (183 countries).

r = 0.84WHO and IHME agree on measuring suicide / self-harm (Pearson)
0.75 → 0.19Cross-validated R² drops once the circular self-harm feature is removed — most apparent skill was leakage
Life expectancy & alcoholStrongest independent drivers of the national suicide rate (SHAP)

Predict a country

A leakage-free model estimates the age-standardized suicide rate (per 100k) with a 90% conformal interval. Runs entirely in your browser — predictions are baked in from the trained model.

Explore every version

Four progressive dashboards, from static visuals to an interactive risk estimator. Each opens in a new tab.

Real data

v0 · Visual gallery

Static, high-variety visuals straight from the raw WHO & IHME data.

Open dashboard →Real data

v1 · Main dashboard

Real-data BI dashboard, ML baseline, and the leakage-free enriched model.

Open dashboard →Synthetic

v2 · Advanced analytics

Methods showcase: clustering, forecasting, graphs, explainability.

Open dashboard →Synthetic

v3 · Risk estimator

Interactive probability tool with calibration and what-if scenarios.

Open dashboard →

Under the hood

Modern, honest methods — with baselines kept visible so nothing is oversold.

SHAP — low life expectancy and high alcohol consumption push predicted risk up.
SHAP — low life expectancy and high alcohol consumption push predicted risk up.
UMAP — countries mapped by mental-health profile, with a suicide-rate gradient.
UMAP — countries mapped by mental-health profile, with a suicide-rate gradient.
Similarity network — 10 communities (modularity 0.72) with bridge countries.
Similarity network — 10 communities (modularity 0.72) with bridge countries.
Cross-source check — WHO vs IHME agreement and measurement difference.
Cross-source check — WHO vs IHME agreement and measurement difference.

How it's built

Machine learning

  • LightGBM + Optuna, nested CV
  • SHAP explainability
  • Conformal prediction intervals
  • Hierarchical mixed-effects (ICC)

Data mining

  • UMAP embedding
  • Country similarity network
  • Subgroup discovery
  • FP-Growth association rules

Engineering

  • Reproducible pipelines
  • FastAPI prediction service
  • Docker + docker-compose
  • CI, tests, LaTeX report