Superset × Claude MCP — Healthcare ED Dashboard

An AI-automated BI dashboard: Apache Superset + Claude MCP builds charts, configures RLS, and deploys — all from natural language.

---

What This Is

A complete Healthcare Emergency Department dashboard built entirely through AI automation. No manual chart creation, no clicking through UI wizards — Claude Code talks to Superset via MCP to:

1. Create datasets from PostgreSQL tables
2. Build 12+ charts (bar, line, heatmap, gauge, pie)
3. Assemble dashboard layouts with filters and cross-links
4. Configure Row-Level Security so each user sees only their authorised data
5. Export everything as code (JSON + Git versioned)

Architecture

┌─────────────────────┐
│   PostgreSQL 16     │  200K+ synthetic ED records
│   (healthcare_db)   │  Star schema: fact_ed_visits + 4 dims
└────────┬────────────┘
         │ SQLAlchemy
┌────────▼────────────┐
│   Apache Superset   │  Dashboard engine
│   (Docker Compose)  │  Native RLS, RBAC, 60+ chart types
└────────┬────────────┘
         │ MCP Server (built-in, Superset 5.0+)
┌────────▼────────────┐
│   Claude Code /     │  Natural language → API calls
│   Claude Desktop    │  "Create a bar chart of wait times
└─────────────────────┘   by triage category from ed_visits"

Dashboard Pages

ED Performance Overview

KPI cards (presentations, ALOS, 4-hour rule %), monthly trend lines, LHD comparison bars.

Triage & Wait Times

Wait time distributions by triage category, hourly presentation heatmaps, breach rate gauges.

Capacity & Flow

Bed occupancy by ward type, patient flow Sankey diagrams, time-to-admission tracking.

Preview

Row-Level Security

Three roles with automatic data filtering:

Role	Sees	Filter
Admin	All LHDs, all wards	No filter
LHD Manager	Their LHD only	lhd_name = {{current_user.lhd}}
Ward Nurse	Their ward only	ward_id = {{current_user.ward}}

RLS rules are configured programmatically via Superset's REST API (8 endpoints), not through the UI.

RLS in Action

Admin sees all data → LHD Manager sees Sydney only → Ward Nurse sees their ward only

Colour Palette

Role	HEX	Usage
Primary	#002664	Main KPIs, chart series 1
Secondary	#146cfd	Chart series 2, emphasis
Teal	#2e808e	Chart series 3, categories
Baby Blue	#8ce0ff	Chart series 4, backgrounds
Soft Pink	#ffb8c1	Accent, light warnings
Alert	#630019	Threshold breaches
Neutral	#d1eeea	Gridlines, backgrounds

Quick Start

git clone https://github.com/yujiyamane/superset-claude-health.git
cd superset-claude-health

# Start Superset + PostgreSQL
docker compose up -d

# Load healthcare data
python scripts/load_data.py

# Connect Claude Desktop via MCP (see docs/mcp_setup.md)
# Then ask Claude: "Create the ED Performance dashboard"

Repository Structure

├── README.md
├── docker-compose.yml          # Superset + PostgreSQL + MCP server
├── superset/
│   ├── superset_config.py      # Superset configuration + MCP + RLS
│   ├── dashboards/             # Exported dashboard JSON (Git-versioned)
│   └── datasets/               # Dataset definitions
├── data/
│   ├── generate_data.py        # Synthetic ED data generator (same as Case 2)
│   └── schema.sql              # PostgreSQL schema (star schema)
├── scripts/
│   ├── load_data.py            # Data loader: CSV → PostgreSQL
│   ├── setup_rls.py            # Programmatic RLS configuration via API
│   ├── setup_roles.py          # RBAC role creation via API
│   └── apply_theme.py          # Colour palette application via API
├── tests/
│   ├── test_data_integrity.py  # Data quality checks
│   ├── test_api_endpoints.py   # Superset API availability
│   ├── test_rls_rules.py       # RLS filtering validation
│   ├── test_colour_theme.py    # Palette consistency
│   └── test_dashboard_export.py # Dashboard-as-Code validation
├── docs/
│   ├── architecture.md         # System architecture
│   ├── mcp_setup.md            # Claude MCP connection guide
│   ├── rls_guide.md            # RLS implementation details
│   └── before_after.md         # Manual vs AI-automated comparison
└── screenshots/
    ├── overview_top.png
    ├── overview_bottom.png
    └── rls_demo.gif

Key Metrics

What	Manual BI Build	AI-Automated (This Project)
Dashboard creation	2-3 days	~30 minutes
RLS configuration	2-4 hours	~5 minutes (scripted)
Theme customisation	1-2 hours	~2 minutes (API)
Reproducibility	Manual screenshots	docker compose up + Git

Tech Stack

Layer	Tool
Data Store	PostgreSQL 16
BI Platform	Apache Superset 5.0+
AI Integration	Claude MCP (Model Context Protocol)
Data Generation	Python (pandas, numpy, Faker)
Orchestration	Docker Compose
Version Control	Git + GitHub
Testing	pytest

---

Dashboard 2: Meeting Room Utilisation

Analysis of meeting room booking patterns across 8 floor levels and 50+ rooms.

Key Metrics

• 50,000+ bookings analysed (2024–2026)
• ~52 minute average booking duration
• 13 rooms most active per day on average
• Peak hours: 9–11 AM and 1–2 PM

Charts

• 4 KPI cards (Total Bookings, Hours Booked, Avg Duration, Utilisation Rate)
• 3 Donut charts (Floor Level, Time of Day, Day of Week distributions)
• Heatmap (Peak booking times by hour × day)
• 3 Bar charts (Floor level, Room name, Utilisation Rate comparisons)
• Combo chart (Bookings vs Avg Duration by Day of Week)
• Time series area (Daily booking volume 2024–2026)
• 6 Native filters (Day, Hour, Organizer, Subject, Room, Floor Level)

Data Generation

python data/generate_meeting_data.py
python scripts/load_meeting_data.py
python scripts/create_meeting_charts.py
python scripts/create_meeting_dashboard.py

---

Related Projects

Case	Focus	Repo
Case 1	Power BI × Healthcare KPI	powerbi-claude-health
Case 2	Streamlit × ED Performance	streamlit-claude-health
Case 3	Life OS × Personal Analytics	lifeos-claude-dashboard
Case 4	Superset × Claude MCP × RLS	This repo

Author

Yuji Yamane — BI Developer @ NSW Health | AI-augmented analytics