mkieffer/Medicaid-Provider-Spending

Medicaid Provider Spending

This dataset packages:

• the raw Medicaid Provider Spending dataset (as published on HHS Open Data), and
• the derived identifier-resolution artifacts produced by this repository when extracting and resolving the unique NPIs and HCPCS/CPT codes found in the raw spending data.

Source Data

• Medicaid Provider Spending (public dataset published by HHS/Open Data): opendata.hhs.gov/datasets/medicaid-provider-spending
• CMS NPPES bulk dissemination files + CMS NPI Registry API (for NPI/provider metadata)
• NLM Clinical Tables HCPCS API (for HCPCS/CPT code metadata)

Source Repository

Project code, build scripts, and pipeline logic live at: ToadResearch/Medicaid_Provider_Spending.

Data Splits

These are exposed on Hugging Face as three dataset configurations (spending, npi, hcpcs) because each has a different schema (the Hub viewer and datasets expect a consistent schema within a single config).

split	contents
spending	Raw Medicaid Provider Spending parquet (as downloaded; unmodified) from HHS Open Data.
npi	Derived NPI resolution artifacts: one row per unique NPI in the spending dataset (deduped). Populated primarily from bulk NPPES dissemination files; the CMS NPI Registry API is used only as fallback for bulk misses. Includes provenance (URL/params), error message, and a raw-or-synthetic JSON payload matching the NPI Registry API response shape.
hcpcs	Derived HCPCS/CPT resolution artifacts: one row per unique code in the spending dataset (deduped). Populated from the resolved mapping cache (Clinical Tables API and/or local CPT/HCPCS fallback). Includes provenance (URL/params), error message, and a raw-or-synthetic JSON payload matching the Clinical Tables response shape.

Usage

from datasets import load_dataset

spending = load_dataset("mkieffer/Medicaid-Provider-Spending", "spending")["spending"]
npi = load_dataset("mkieffer/Medicaid-Provider-Spending", "npi")["npi"]
hcpcs = load_dataset("mkieffer/Medicaid-Provider-Spending", "hcpcs")["hcpcs"]

Split Schemas

spending

column	type	description
BILLING_PROVIDER_NPI_NUM	string	Billing provider NPI
SERVICING_PROVIDER_NPI_NUM	string	Servicing provider NPI
HCPCS_CODE	string	HCPCS/CPT code
CLAIM_FROM_MONTH	string	Claim month (YYYY-MM-01)
TOTAL_UNIQUE_BENEFICIARIES	int64	Unique beneficiaries
TOTAL_CLAIMS	int64	Total claims
TOTAL_PAID	float64	Total paid (USD)

npi

All columns in this split are string (some are JSON-encoded strings).

column	type	description
npi	string	NPI
basic	string (JSON)	results[0].basic
addresses	string (JSON)	results[0].addresses
practice_locations	string (JSON)	results[0].practiceLocations
taxonomies	string (JSON)	results[0].taxonomies
identifiers	string (JSON)	results[0].identifiers
other_names	string (JSON)	results[0].other_names
endpoints	string (JSON)	results[0].endpoints
url	string	Request URL when an API call happened; otherwise a bulk sentinel like nppes_bulk:*
error_message	string (nullable)	Error message if lookup failed
api_run_id	string	Local pipeline run id
requested_at_utc	string	Request timestamp (UTC) or bulk-export generation timestamp
request_params	string (JSON)	Request params / provenance captured by the pipeline
results	string (JSON)	Full results array
response_json	string (JSON)	Full raw API JSON payload when an API call happened; otherwise a synthetic payload in the same shape

hcpcs

All columns in this split are string (some are JSON-encoded strings).

column	type	description
hcpcs_code	string	HCPCS/CPT code
ef_short_desc	string (JSON)	extra_fields.short_desc (list)
ef_long_desc	string (JSON)	extra_fields.long_desc (list)
ef_add_dt	string (JSON)	extra_fields.add_dt (list)
ef_act_eff_dt	string (JSON)	extra_fields.act_eff_dt (list)
ef_term_dt	string (JSON)	extra_fields.term_dt (list)
ef_obsolete	string (JSON)	extra_fields.obsolete (list)
ef_is_noc	string (JSON)	extra_fields.is_noc (list)
response_codes	string (JSON)	Codes array returned by the API
response_display	string (JSON)	Display array returned by the API
response_extra_fields	string (JSON)	Extra fields object returned by the API
url	string	Request URL when an API call happened; otherwise a cache/fallback sentinel like hcpcs_cache:*
error_message	string (nullable)	Error message if lookup failed
api_run_id	string	Local pipeline run id
requested_at_utc	string	Request timestamp (UTC) or cache-export generation timestamp
request_params	string (JSON)	Request params / provenance captured by the pipeline
response_json	string (JSON)	Full raw API JSON payload when an API call happened; otherwise a synthetic payload in the same shape

Parquet Null / Empty-List Audit

Auto-generated by ./build_datasets.sh (or cargo run --release --manifest-path build_datasets/Cargo.toml -- --null-check).

• Generated at (unix seconds): 1771173117
• NPI parquet: data/output/npi.parquet
• HCPCS parquet: data/output/hcpcs.parquet

Notes:

• null_count counts actual Parquet nulls.
• empty_list_count counts the literal string value "[]" (JSON-encoded empty list).

NPI (data/output/npi.parquet)

column	rows_total	null_count	null_pct	empty_list_count	empty_list_pct
error_message	1802136	1799569	99.86%	0	0.00%
other_names	1802136	4176	0.23%	1657485	91.97%
endpoints	1802136	4176	0.23%	1591727	88.32%
practice_locations	1802136	4176	0.23%	1480277	82.14%
identifiers	1802136	4176	0.23%	1208848	67.08%
addresses	1802136	4176	0.23%	20097	1.12%
taxonomies	1802136	4176	0.23%	20097	1.12%
basic	1802136	4176	0.23%	0	0.00%
results	1802136	4065	0.23%	111	0.01%
response_json	1802136	2567	0.14%	0	0.00%
api_run_id	1802136	0	0.00%	0	0.00%
npi	1802136	0	0.00%	0	0.00%
request_params	1802136	0	0.00%	0	0.00%
requested_at_utc	1802136	0	0.00%	0	0.00%
url	1802136	0	0.00%	0	0.00%

HCPCS (data/output/hcpcs.parquet)

column	rows_total	null_count	null_pct	empty_list_count	empty_list_pct
error_message	10881	6271	57.63%	0	0.00%
ef_act_eff_dt	10881	0	0.00%	4610	42.37%
ef_add_dt	10881	0	0.00%	4610	42.37%
ef_is_noc	10881	0	0.00%	4610	42.37%
ef_long_desc	10881	0	0.00%	4610	42.37%
ef_obsolete	10881	0	0.00%	4610	42.37%
ef_short_desc	10881	0	0.00%	4610	42.37%
ef_term_dt	10881	0	0.00%	4610	42.37%
response_codes	10881	0	0.00%	4610	42.37%
response_display	10881	0	0.00%	4610	42.37%
api_run_id	10881	0	0.00%	0	0.00%
hcpcs_code	10881	0	0.00%	0	0.00%
request_params	10881	0	0.00%	0	0.00%
requested_at_utc	10881	0	0.00%	0	0.00%
response_extra_fields	10881	0	0.00%	0	0.00%
response_json	10881	0	0.00%	0	0.00%
url	10881	0	0.00%	0	0.00%

Unmapped / Unresolved Identifier Counts

For the build that produced this dataset:

• NPI: 24,272 unique NPIs were not successfully mapped/resolved (15,012 request errors; 9,260 not_found).

  • Many of these were invalid NPI identifiers (4,111 unique): 2,928 non-numeric, 971 wrong-length, 205 checksum-invalid (Luhn), and 7 placeholders.

  • Some NPIs have been inactive/deactivated in NPPES/NPI Registry (status can change over time); when present, check the registry payload for an active/inactive status (and/or NPPES deactivation fields) before treating an NPI as current.

  • Full format breakdown from triage:

    inferred type	count	%	meaning
    npi_luhn_valid	20,161	83.06%	looks like a well-formed NPI (10 digits + valid check digit)
    non_numeric	2,928	12.06%	contains non-digits (often needs upstream cleaning)
    numeric_wrong_len	971	4.00%	all digits but not 10 digits long
    npi_luhn_invalid	205	0.84%	10 digits, but invalid Luhn check digit
    placeholder_or_invalid	7	0.03%	placeholder-like values (e.g., all zeros)

• HCPCS/CPT: 4,610 unique codes were not successfully mapped/resolved (4,608 not_found; 2 request errors).

  • Full resolved/unresolved pattern breakdown from triage:

    pattern family	count	%
    CPT_or_HCPCS_L1_5digit (5-digit CPT/HCPCS)	2,169	47.05%
    HCPCS_level_II (A0000 style)	474	10.28%
    CDT (D####)	449	9.74%
    CPT_category_II (####F)	358	7.77%
    alphanum_5char_unknown	310	6.72%
    revenue_code_4digit	272	5.90%
    4digit_plus_letter_other	124	2.69%
    CPT_PLA (####U)	110	2.39%
    CPT_category_III (####T)	96	2.08%
    word_or_flag	78	1.69%
    unknown	57	1.24%
    HCPCS_L2_plus_modifier	39	0.85%
    CPT_5digit_plus_modifier	24	0.52%
    drg_like_3digit	15	0.33%
    icd10pcs_like_7char	11	0.24%
    modifier_2char	7	0.15%
    numeric_6to8_unknown	7	0.15%
    placeholder_or_invalid	6	0.13%
    CDT_plus_suffix	4	0.09%

  • 458 unique codes were flagged by triage as "needs review" (e.g., flags/words, placeholders, or unknown alphanumeric patterns).

  • The 2 request errors are included in unknown (for example H)015, NT/WKN).