VaidhyaMegha/mental-health-kg

Mental Health Knowledge Graph

113,710 nodes. 1,665,153 edges. US behavioural-health provision as a graph: which facilities exist and what they offer, which clinicians are licensed to practise, where the federal government designates a shortage, and a simulated population to measure coverage against.

Built with Samyama Graph. Loader and ETL: samyama-ai/mental-health-kg.

Built for referral routing — which help exists where, for whom, in what language, at what price — and for coverage analytics, the question of where that help does not exist.

⚠️ Read this first: people in this graph

No real patient data. The 3,500 Patient nodes and 11 Condition nodes are synthetic, generated by Synthea (MITRE, Apache 2.0). They carry synthetic: true and source: "Synthea". No survivor, no session, no transcript, no contact record.

Real clinicians do appear, from the NPPES NPI Registry — a public federal register that CMS publishes in full. Provider nodes carry name, NPI, business practice city/state/ZIP, business phone and taxonomy.

Professional licence numbers have been removed from this dataset. The upstream snapshot carries licence and licence_state on Provider, populated for 41,485 of them. Both columns were dropped before publishing, honouring the commitment in the source repository's own dataset card that "licence numbers are masked in any published demo artifact."

For the same reason no .sgsnap snapshot is shipped here — the snapshot still contains those licence numbers, so publishing it would undo the masking. Rebuild from the ETL if you need the full graph.

All remaining Provider fields are business-practice details from a public federal register, not personal contact details.

Sources — all redistributable

Source	Publisher	Licence	Produces
FindTreatment.gov	SAMHSA / BHSIS	US federal government work — public domain	Facility, Service, ServiceCategory, Language, FacilityType, State
NPPES NPI Registry	CMS	US federal government work — public domain	Provider, Taxonomy
HRSA HPSA	HRSA / HHS	US federal government work — public domain	ShortageArea, County
Synthea	The MITRE Corporation	Apache 2.0	Patient, Condition

The NUCC taxonomy code set supplies the readable classification on Taxonomy nodes — a lookup over the NPPES layer rather than a fifth source.

Nothing here restricts redistribution. Three sources are US federal government works and therefore outside copyright; Synthea is Apache 2.0. license: other is used because no single SPDX tag covers "public domain plus Apache-2.0", not because the terms are unclear.

The graph carries its own provenance: nodes/datasource.csv holds one row per source with its fetch date, refresh cadence, caveats and a real_data flag.

Files

Nodes carry an id; edges reference those ids as src and tgt.

nodes/

File	Rows	Real?
provider.csv	82,978	real — 74,562 individuals, 8,416 organisations
facility.csv	17,254	real
shortagearea.csv	6,420	real
patient.csv	3,500	synthetic
county.csv	3,043	real
service.csv	313	real
taxonomy.csv	76	real
state.csv	52	real
servicecategory.csv	33	real
language.csv	24	real
condition.csv	11	synthetic
datasource.csv	4	provenance manifest
facilitytype.csv	2	real

edges/

File	Rows	Connects
offers.csv	1,417,479	Facility → Service
has_taxonomy.csv	82,978	Provider → Taxonomy
practices_in.csv	82,978	Provider → State
has_type.csv	23,293	Facility → FacilityType
located_in.csv	17,254	Facility → State
speaks.csv	15,740	Facility → Language
covers.csv	7,842	ShortageArea → County
has_condition.csv	3,918	Patient → Condition
in_county.csv	3,500	Patient → County
lives_in.csv	3,500	Patient → State
has_providers.csv	3,329	State → Taxonomy
in_state.csv	3,029	County → State
in_category.csv	313	Service → ServiceCategory

OFFERS is 85% of the graph — that is the referral surface.

Usage

from datasets import load_dataset

facilities = load_dataset("VaidhyaMegha/mental-health-kg", "facility", revision="v1.0")

The question this graph was built to answer — where can a survivor of intimate partner violence get trauma counselling, in Spanish, on a sliding fee scale? — is a multi-way intersection over one node, which is exactly what a graph does well:

MATCH (f:Facility)-[:OFFERS]->(a:Service),
      (f)-[:OFFERS]->(b:Service),
      (f)-[:SPEAKS]->(l:Language)
WHERE a.name CONTAINS 'Trauma' AND b.name CONTAINS 'Sliding'
  AND l.name = 'Spanish'
RETURN f.name, f.city, f.state, f.phone

// Coverage gap: where does designated shortage meet thin provider supply?
MATCH (s:ShortageArea)-[:COVERS]->(c:County)-[:IN_STATE]->(st:State)
OPTIONAL MATCH (p:Provider)-[:PRACTICES_IN]->(st)
RETURN st.name, count(DISTINCT c) AS shortage_counties,
       count(DISTINCT p) AS providers
ORDER BY shortage_counties DESC

Verification

• Totals reconcile against the snapshot header: 113,710 nodes, 1,665,153 edges.
• 0 dangling edges — every src and tgt resolves to a node in this dataset.
• 21 orphan nodes, all explainable: 4 DataSource (a provenance manifest, connected to nothing by design), 12 ShortageArea with no county link, 5 Taxonomy with no provider.
• Label counts match the source repository's own dataset card exactly, all thirteen.
• Licence columns confirmed absent from every published file.

Limitations

• A snapshot of 2026-08-18, not a live directory. Facilities close, clinicians move, HPSA designations are revised. Do not use it to tell someone a service exists today — verify against FindTreatment.gov first.
• NPPES tells you who is licensed, not who is accepting patients, nor where they actually work day to day. It is a registry, not a directory of availability.
• Provider coverage is MA and VT, not national. Facilities and shortage areas are national.
• The synthetic layer is Synthea's, with Synthea's demographics. It measures coverage geometry; it is not a claim about any real population's need. Any figure derived from it carries that caveat.
• Services are facility self-reported categories, with the inconsistency that implies.

Not clinical, legal or referral advice. For someone who needs help now, use FindTreatment.gov or call/text 988 (US).

Citation

Mental Health Knowledge Graph, v1.0 (113,710 nodes, 1,665,153 edges).
Built with Samyama Graph. https://huggingface.co/datasets/VaidhyaMegha/mental-health-kg
ETL: https://github.com/samyama-ai/mental-health-kg
Sources: FindTreatment.gov (SAMHSA), NPPES NPI Registry (CMS), HRSA HPSA — US federal
         government works, public domain; Synthea (MITRE), Apache 2.0.
Snapshot: 2026-08-18. Professional licence numbers removed.