data trove cms hospital database

Overview

Source: /home/coolhand/html/datavis/data_trove/data/healthcare/cms_hospitals_2025.csv

Saturn profiled 5,421 rows across 38 columns. The stats below are deterministic and machine-readable; the prose is a language-model interpretation of those stats (opt-in, added after the fact, never sees raw rows).

Facility ID text identifier

This column is a fixed-length, zero-padded 6-character alphanumeric facility identifier — likely a government or regulatory facility code (e.g., CMS Facility ID or similar). Every one of the 5,421 rows has a distinct value with zero nulls and zero duplicates, confirming it functions as a primary key. All values are exactly 6 characters long (min, mean, max = 6) and fully uppercase, consistent with a structured code scheme. The top visible values follow a numeric sequence starting with '010001', suggesting geographic or administrative ordering.

Treatment: Retain as a primary key for row identification and use as a join key to link facility-level metadata tables.

Facility Name text label

This column contains healthcare facility names — overwhelmingly hospitals and medical centers, as evidenced by 'hospital' appearing 2,740 times and 'medical'/'health'/'center' dominating the top-word list across 5,421 rows. Nearly all values are ALL-CAPS (99.3%), suggesting a standardized administrative data entry convention. The near-unique alert is noteworthy: with 5,286 unique values out of 5,421 rows (duplicate_rate 2.5%, 135 duplicates), some facilities appear more than once, which could indicate multiple records per facility rather than a clean one-row-per-facility structure. Average name length of ~29 characters and ~4 words per name is consistent with formal institutional naming patterns.

Treatment: Use as a human-readable label or grouping key; normalize case before display; investigate 135 duplicate entries for potential record linkage issues before modelling.

Address text metadata

This column contains physical street addresses, confirmed by dominant top words: 'street' (1,036), 'avenue' (580), 'drive' (511), 'road' (507), and directional prefixes like 'north', 'east', 'west', 'south'. Nearly all values are uppercased (99.2% allcaps_rate), consistent with address data sourced from a government registry or standardised mailing system. With 5,387 unique values out of 5,421 rows and only 34 duplicates (0.6% duplicate rate), the column is near-unique — the small number of duplicates may indicate shared addresses such as apartment buildings or data entry issues worth investigating.

Treatment: Do not use as a predictive feature directly; parse into structured components (number, street, direction, suffix) or geocode for spatial modelling.

City/Town text label

This column contains US city and town names, stored in ALL-CAPS format (99.4% allcaps rate), consistent with a standardized postal or administrative data source. With 3,049 unique values across 5,421 rows, the duplicate rate of 43.8% (2,372 duplicates) is expected for a geographic field — major cities like CHICAGO (34) and HOUSTON (31) naturally repeat. The near-complete one-word rate (77.1%) alongside multi-word entries like 'OKLAHOMA CITY' and 'LOS ANGELES' is normal for city names. No nulls are present, indicating good completeness.

Treatment: Standardize casing, then use as a grouping/aggregation key or encode as a high-cardinality categorical feature.

State categorical feature

This column contains US state abbreviations, covering 56 distinct values (50 states plus likely DC and US territories). The distribution is notably spread — entropy ratio of 0.917 indicates near-uniform coverage — but TX leads with 462 occurrences (8.5% of 5,421 rows), followed by CA (378) and FL (221), reflecting population-weighted representation. The cardinality of 56 slightly exceeding 50 states warrants a quick audit to confirm no invalid or duplicate codes exist.

Treatment: One-hot encode or use target encoding depending on model type; verify the 6 non-standard codes (beyond 50 states) are valid territories or DC.

ZIP Code numeric feature

This column contains US postal (ZIP) codes stored as integers rather than strings, covering the full national range from 603 (Puerto Rico/Northeast) to 99929 (Alaska), with 4721 distinct values across 5421 rows. The numeric treatment is problematic: leading zeros are silently dropped (the minimum of 603 almost certainly represents 00603), making string-based joins or lookups unreliable. The near-flat kurtosis (−0.99) and low skew (−0.16) indicate fairly broad geographic spread with no dominant regional cluster, which is notable given the wide IQR of 43333.

Treatment: Cast to zero-padded 5-character string before any join, geocoding, or feature engineering to restore lost leading zeros.

County/Parish text label

This column contains US county (or parish) names, stored entirely in uppercase (allcaps_rate 1.0), consistent with a standardised geographic reference field. With 1,555 unique values across 5,421 rows and a duplicate_rate of 0.7132, the same county names recur frequently — expected given that many records share common counties like LOS ANGELES (88), JEFFERSON (59), and COOK (59). The high one_word_rate (0.873) reflects that most US county names are single tokens, while multi-word entries like 'LOS ANGELES' and 'SAN ...' account for the remainder.

Treatment: Encode as a categorical geographic grouping variable; consider joining to a FIPS code lookup for spatial analysis.

Telephone Number text feature

This column contains North American telephone numbers, all formatted uniformly at exactly 14 characters (consistent with the '(NXX) NXX-XXXX' pattern) with zero nulls across 5,421 records. The allcaps_rate of 1.0 is an artifact of digit/punctuation-only strings rather than alphabetic content. Surprisingly, 38 duplicate phone numbers exist (duplicate_rate ≈ 0.007), meaning distinct records share the same telephone number, which warrants investigation for data quality issues. Top area codes such as (406), (605), and (402) suggest a predominantly rural US geographic distribution.

Treatment: Flag 38 duplicate values for deduplication review; treat as a categorical or string identifier — do not embed numerically.

Hospital Type categorical label

This column classifies each hospital record into one of 8 facility-type categories, making it a standard categorical label for hospital segmentation. 'Acute Care Hospitals' dominates at 57.6% of records (3,120 of 5,421), creating notable class imbalance. The extreme tail is striking: 'Long-term' hospitals appear only 4 times, and 'Acute Care - Department of Defense' just 32 times, meaning minority classes may be statistically unreliable for subgroup analysis. There are no nulls and no unexpected values.

Treatment: One-hot encode for modelling; be cautious with rare classes ('Long-term' n=4, 'DoD' n=32) which may need grouping or exclusion.

Hospital Ownership categorical feature

This column captures the legal/operational ownership classification of hospitals, with 12 distinct categories across 5,421 records and no nulls. The dominant category is 'Voluntary non-profit - Private' at 42.3% (2,291 records), creating notable imbalance — the top value alone accounts for more than twice the second-largest category ('Proprietary' at 1,067). The entropy ratio of 0.72 indicates moderate but uneven spread across categories, with minority classes like 'Government - Federal' (44) and 'Physician' (74) being quite sparse.

Treatment: One-hot encode or target-encode for modelling, but consider grouping sparse categories (e.g., 'Government - Federal' with n=44, 'Physician' with n=74) to avoid instability.

Emergency Services categorical feature

This column is a binary flag indicating whether emergency services are present or available for a given record. The dominant value is 'Yes' at 83.1% (4505 of 5421 rows), making the distribution notably imbalanced — 'No' accounts for only 916 records (roughly 1-in-6). No nulls exist, and entropy of 0.655 reflects the skew away from a balanced 50/50 split.

Treatment: Encode as binary (1/0); note class imbalance (83/17 split) and apply appropriate weighting or stratified sampling if used as a target or predictor.

Meets criteria for birthing friendly designation categorical label

This column flags whether a facility meets criteria for a 'birthing friendly' designation, and every non-null value is 'Y' (top_rate = 1.0, cardinality = 1). That means it carries zero discriminative information among observed records — there are no 'N' values at all. More striking, 58.24% of rows are null, leaving only 2,264 usable observations out of 5,421; the nulls likely represent facilities that were not assessed or do not offer birthing services.

Treatment: Treat non-null as a binary indicator of assessed-and-qualifying facilities; nulls should be coded as a distinct category (e.g. 'not assessed') rather than imputed, as they almost certainly reflect structural missingness rather than random dropout.

Hospital overall rating categorical label

This column represents a CMS-style 1–5 star overall hospital rating, stored as a categorical field with 6 distinct values. The most striking finding is that 47.1% of all 5,421 rows carry the value 'Not Available', making missing ratings the single largest 'category'—outnumbering any numeric rating tier. Among hospitals that do have a rating, scores cluster around 3 and 4 (937 and 765 occurrences respectively), with the extremes (1 and 5) being relatively rare.

Treatment: Convert numeric strings to ordinal integers; treat 'Not Available' as a distinct missing indicator (do not encode as a numeric level) and consider imputation or a separate missingness flag before modelling.

Hospital overall rating footnote categorical metadata

This column contains footnote codes attached to hospital overall ratings, where each numeric code references a specific explanatory note (e.g., data suppression reason, insufficient data, or special methodology). The 52.7% null rate is expected since most hospitals with valid ratings require no footnote. However, among non-null rows, value '16' dominates at 65.4% of present values (1,676 occurrences), suggesting a single suppression or caveat reason is overwhelmingly common; the presence of a compound value '16, 23' indicates some records carry multiple footnote codes, which could complicate downstream parsing.

Treatment: Treat nulls as 'no footnote'; split multi-code entries on comma and one-hot encode each footnote code separately before modelling.

MORT Group Measure Count categorical feature

This column records the count of MORT (mortality) group measures evaluated per entity, with only two distinct values across 5,421 rows: '7' (dominant, 84.1%) and 'Not Available' (15.9%). The cardinality of 2 and the presence of 'Not Available' as the only alternative to '7' suggests this is effectively a binary flag — entities either have a full complement of 7 mortality measures or are excluded from measurement entirely. The 'Not Available' string rather than a true null (null_rate is 0.0) means missingness is encoded as a sentinel value and must be handled explicitly.

Treatment: Recode 'Not Available' to null or a binary indicator before modelling; if '7' is constant for valid records, consider dropping or using only the missingness flag.

Count of Facility MORT Measures categorical feature

This column records how many mortality quality measures are reported for a given facility, taking integer values 1–7 plus a sentinel string 'Not Available'. Despite being numeric in nature, it was parsed as categorical, likely because of that mixed-type sentinel. Surprisingly, 'Not Available' is the single most frequent value at 32.8% (1,777 of 5,421 rows), meaning roughly a third of facilities have no mortality measure count on record. The remaining values are fairly evenly spread across 1–7, suggesting the missingness is not random but tied to whether a facility participates in mortality reporting at all.

Treatment: Separate 'Not Available' into a boolean missingness indicator, then cast the remaining values to integer for use as an ordinal or numeric feature.

Count of MORT Measures Better categorical feature

This column encodes how many mortality (MORT) measures a hospital performs better than the national benchmark, stored as a categorical string despite being an ordinal count ranging from 0 to 7. The dominant value is '0' (3,136 rows, 57.8%), meaning most hospitals beat no mortality benchmarks, while 1,777 rows (32.8%) are 'Not Available', leaving only ~9.4% of hospitals showing any above-average mortality performance. The absence of value '6' except for a single record and the jump from '5' to '7' (skipping '6' effectively) signals a highly right-skewed, near-zero distribution with a substantial missingness class masquerading as a category.

Treatment: Split 'Not Available' into a binary missing indicator, then cast remaining values to integer for ordinal or numeric modelling.

Count of MORT Measures No Different categorical feature

This column represents a count of mortality (MORT) measures where a hospital's performance was rated 'No Different' from the national benchmark, stored as a categorical string rather than a numeric type. The most striking signal is that 32.8% of rows (1,777 of 5,421) carry 'Not Available', effectively a masked null despite the reported null_rate of 0.0. The numeric values range from 0 to 7, with '0' being extremely rare (only 12 occurrences), suggesting nearly all reporting hospitals have at least one mortality measure in this category.

Treatment: Recode 'Not Available' as a true null or a separate indicator flag, then cast remaining values to integer for modelling.

Count of MORT Measures Worse categorical feature

This column counts how many mortality (MORT) quality measures a hospital scored 'worse than expected' on, stored as a categorical string rather than an integer. The dominant value is '0' (3,266 rows, 60.2%), indicating most hospitals have no worse-than-expected mortality flags, but 1,777 rows (32.8%) carry 'Not Available', which is a substantial missingness proxy masked by a zero null_rate. Only 378 hospitals have one or more worse measures, and the counts are right-skewed, topping out at 5.

Treatment: Convert numeric strings to integers, recode 'Not Available' as NaN, then treat as an ordinal count feature; consider a binary flag for any worse measure given extreme skew toward 0.

MORT Group Footnote numeric label

This column contains footnote codes associated with a mortality (MORT) group, encoded as numeric values despite functioning as a categorical label. With only 4 unique values (5.0, and likely 19.0, 23.0, and one other near the mean) drawn from a set bounded by 5–23, this is clearly a small enumeration of footnote categories rather than a true numeric measure. The 67.2% null rate is the dominant signal — meaning footnotes apply to fewer than a third of rows, suggesting they flag exceptional or conditional cases. The near-flat kurtosis (−1.96) and bimodal-like IQR of 14.0 confirm the values are widely spread across only a few discrete codes.

Treatment: Cast to categorical/string; treat nulls as 'no footnote' class; one-hot encode if used as a feature.

Safety Group Measure Count categorical feature

This column represents a count of safety group measures, but is typed as categorical with only two distinct values: '8' (a fixed numeric string) and 'Not Available'. The dominance of a single value '8' across 84.1% of 5,421 rows (4,558 occurrences) suggests this may be a schema-constrained field or a denormalized count that rarely varies. The presence of 'Not Available' as the only alternative (863 rows, ~15.9%) rather than a true numeric null is surprising and indicates missing data was encoded as a string sentinel rather than left null.

Treatment: Binarize into a flag (count_available vs. not_available); treat 'Not Available' as missing before any numeric use.

Count of Facility Safety Measures categorical feature

This column represents an ordinal count of safety measures implemented at a facility, stored as a categorical/string type with values ranging from 1 to 8. The most striking signal is that 38.1% of the 5,421 rows (2,065 records) carry the value 'Not Available', which is the single largest category and likely reflects missing or unreported data rather than a meaningful category. The numeric values show a non-uniform distribution, with '7' being the most common true count (733 occurrences) and '4' the rarest (223), suggesting a rough bimodal tendency toward higher counts. The column should be treated as ordinal numeric after separating out 'Not Available' as a distinct missingness indicator.

Treatment: Cast numeric strings to integer, encode 'Not Available' as a separate missingness flag or impute, then use as ordinal feature.

Count of Safety Measures Better categorical feature

This column represents a count of safety measures rated 'better' for a given entity (likely a healthcare facility or similar regulated site), stored as a categorical rather than numeric type. The dominant value is 'Not Available' at 38.1% of rows (2,065 of 5,421), which masks the underlying numeric distribution; among records with actual counts, the distribution is strongly right-skewed, with 1,548 zeros and only 3 records reaching the maximum value of 6. Analysts should be aware that the high 'Not Available' rate introduces substantial missingness risk if this field is coerced to numeric.

Treatment: Separate 'Not Available' into a missingness indicator, then cast remaining values to integer for ordinal or numeric modelling.

Count of Safety Measures No Different categorical feature

This column represents a count of safety measures rated as 'no different' (presumably compared to a benchmark), stored as a categorical type despite containing numeric-looking values 0–8. The dominant value is 'Not Available' at 38.1% of 5,421 rows (2,065 records), which is a significant missingness proxy masking as a category rather than a true null — notable since the null_rate is 0.0. The remaining values follow a rough bell-curve peaking at 5, with very sparse representation at the extremes (8 appears only 10 times, 0 only 20 times).

Treatment: Recode 'Not Available' as NaN, cast remaining values to integer, then treat as ordinal or numeric feature.

Count of Safety Measures Worse categorical feature

This column represents a count of safety measures that have deteriorated, stored as a categorical field with only 5 distinct values (0, 1, 2, 3, and 'Not Available'). The dominant value is '0' at 54.3% of rows (2,941 records), indicating most entities have no worsening safety measures, but 38.1% of rows (2,065) carry 'Not Available' rather than a numeric zero, suggesting a meaningful distinction between 'measured and none found' vs. 'not assessed'. The actual worsening counts (1–3) are rare and sharply drop off, with only 6 records reaching a count of 3.

Treatment: Encode 'Not Available' as a separate indicator flag, then cast numeric strings to integer for ordinal or count-based modelling.

Safety Group Footnote numeric label

This column encodes footnote reference numbers attached to safety groups, with only 4 distinct integer values (5, 19, 23, and at least one other given the IQR and mean) acting as categorical codes rather than true continuous quantities. The 61.8% null rate is a major signal: most records carry no footnote, making non-null values the exception. The platykurtic distribution (kurtosis ≈ −1.81) and near-zero skew confirm a flat, discrete distribution across a handful of categories, not a numeric measurement. Despite being stored as numeric, this should be treated as a sparse categorical indicator.

Treatment: Cast to nullable categorical; one-hot or ordinal encode the 4 distinct values, and treat nulls as a fifth 'no footnote' category.

READM Group Measure Count categorical feature

This column represents the count of readmission group measures associated with a hospital record, stored as a categorical type despite being a numeric concept. It has only two distinct values across 5,421 rows: '11' (84.1% of records) and 'Not Available' (15.9%), with zero nulls. The near-total dominance of a single value '11' and the use of 'Not Available' as a sentinel string (rather than a true null) are both worth flagging — the column carries almost no discriminative power as a feature.

Treatment: Convert 'Not Available' to null, cast to integer, then assess whether near-zero variance warrants dropping before modelling.

Count of Facility READM Measures categorical feature

This column represents the count of readmission measures available for a healthcare facility, stored as a categorical/string type despite containing numeric values. The most striking signal is that 'Not Available' is the dominant value at 21.2% of all 5,421 rows (1,150 records), meaning roughly one-in-five facilities have no reportable readmission measure count. The remaining values span a narrow integer range (at least 2–11 visible), with near-uniform distribution across them given the high entropy ratio of 0.965, suggesting facilities tend to report varying but substantive numbers of measures.

Treatment: Cast numeric strings to integer, encode 'Not Available' as NaN or a separate missingness indicator flag before modelling.

Count of READM Measures Better categorical feature

This column represents a small integer count (0–5) of how many readmission measures a hospital performed 'better' than a national benchmark, stored as a categorical/string field. The dominant value is '0' at 61.5% (3,332 rows), meaning most hospitals beat no readmission benchmarks at all. Notably, 1,150 rows (21.2%) carry 'Not Available' rather than a numeric zero, signalling a data-availability distinction that must be resolved before any numeric analysis.

Treatment: Separate 'Not Available' into a missingness indicator flag, then cast remaining values to integer for ordinal or numeric modelling.

Count of READM Measures No Different categorical feature

This column represents the count of readmission measures rated 'No Different' from the national rate for a given hospital, with valid integer values ranging from 1 to at least 9 (cardinality 13 suggests values up to ~12). The most striking signal is that 'Not Available' is the single most frequent value at 21.2% (1,150 of 5,421 rows), meaning roughly one in five hospitals has no reportable count — this non-numeric sentinel is stored as a string, making the column categorical despite its inherently ordinal numeric nature. Among reportable hospitals, the distribution across counts 1–9 is remarkably flat (372–497 each), suggesting no strong clustering around a typical score. The high entropy ratio (0.92) reflects this near-uniform spread across the 13 distinct values.

Treatment: Separate 'Not Available' into a missingness indicator flag, then cast remaining values to integer for ordinal/numeric modelling.

Count of READM Measures Worse categorical feature

This column records the number of readmission quality measures on which a hospital performed worse than the national benchmark, stored as a categorical/string field despite being fundamentally ordinal-numeric. The dominant value is '0' (2,988 of 5,421 rows, 55.1%), indicating most facilities had no worse-than-average readmission measures, while 'Not Available' accounts for 1,150 rows (21.2%) — a substantial missingness masked by a zero null_rate. The numeric range spans 0–7, but the distribution is heavily right-skewed with only 3 hospitals scoring 6 or higher.

Treatment: Recode 'Not Available' as NaN, cast remaining values to integer, then treat as an ordinal feature or apply count-based encoding before modelling.

READM Group Footnote numeric metadata

This column is a coded footnote indicator for a readmission (READM) metric group, taking only 3 distinct numeric values (5, 19, 22) despite 5,421 rows. The 78.79% null rate is the dominant signal — nulls almost certainly mean 'no footnote applies', making non-null values the exception rather than the rule. The three values are likely lookup codes referencing a footnote legend (e.g., CMS suppression or reliability flags), not a continuous numeric measure, despite being stored as numeric.

Treatment: Treat as categorical; map the 3 codes (5, 19, 22) to their footnote definitions and one-hot encode or use as a suppression/reliability flag.

Pt Exp Group Measure Count categorical feature

This column records the count of patient experience group measures, but is stored as a categorical with only two distinct values: '8' (the actual measure count) and 'Not Available'. The dominant value '8' appears in 84.1% of rows (4,558 of 5,421), while 'Not Available' accounts for the remaining 15.9% (863 rows) — suggesting a structured survey instrument with a fixed number of measures for eligible facilities, and a meaningful missingness pattern for ineligible or non-reporting ones.

Treatment: Treat 'Not Available' as a distinct eligibility/reporting flag; convert '8' to integer if numeric operations are needed, or encode as binary eligible/not-eligible indicator.

Count of Facility Pt Exp Measures categorical feature

This column represents a count of facility patient experience measures, but despite its numeric-sounding name it is stored as a categorical with only 2 distinct values: '8' (58.2% of rows, n=3,154) and 'Not Available' (41.8%, n=2,267). The near-uniform presence of a single numeric value ('8') suggests this count is fixed or standardized across facilities that report it, making the column effectively a binary indicator of data availability rather than a meaningful numeric measure. No nulls exist; missingness is encoded as the string 'Not Available'.

Treatment: Re-encode as binary flag (1 = '8', 0 = 'Not Available') since the numeric value is invariant; drop or use as availability indicator.

Pt Exp Group Footnote numeric label

This column appears to store numeric footnote codes attached to a patient experience group field, with only 3 distinct values (5, 19, and 22 inferred from min/Q1/Q3/max) acting as categorical tags rather than true measurements. The 58.18% null rate is flagged as an alert, meaning footnotes are absent for the majority of records — consistent with footnotes being exception annotations rather than universal attributes. The platykurtic distribution (kurtosis −1.66) and the fact that Q1 equals the median and minimum (all 5.0) confirm the values cluster heavily at the lowest code, with the IQR of 14 spanning to 19.

Treatment: Treat as a low-cardinality categorical; one-hot encode the 3 footnote codes and model nulls as a fourth explicit category ('no footnote').

TE Group Measure Count categorical feature

This column represents a count of TE (likely Test/Treatment Episode) group measures, stored as a categorical rather than numeric type — a likely encoding artefact. It has only two distinct values across 5,421 rows: '12' (84.1% of rows) and 'Not Available' (15.9%), suggesting it is effectively a binary availability flag masquerading as a count. The dominance of a single numeric value '12' with no other numeric variation is surprising and implies the measure count may be fixed/standardised across records rather than truly variable.

Treatment: Convert '12' to numeric and 'Not Available' to null, then treat as a binary missingness indicator or drop if constant after imputation.

Count of Facility TE Measures categorical feature

This column represents the number of Technical Efficiency (TE) measures recorded per facility, stored as a categorical/string type despite being fundamentally numeric with integer values ranging at least from 4 to 12. The most surprising signal is that 'Not Available' is the single most frequent value at 17.1% (928 of 5421 rows), indicating a meaningful data-availability gap that is structurally encoded as a string sentinel rather than a null — zero nulls are reported. Entropy ratio of 0.93 across only 13 categories suggests a near-uniform distribution among the numeric values, with no strong concentration beyond the 'Not Available' category.

Treatment: Convert numeric string values to integer, recode 'Not Available' as NaN, then decide whether to impute or flag as a separate missingness indicator before modelling.

TE Group Footnote numeric metadata

This column encodes footnote reference numbers attached to a 'TE Group' field, taking only 3 distinct numeric values (5, 19, 22) across the entire dataset. 82.88% of rows are null, meaning footnotes apply to a small minority of records. The IQR of 0 and median/Q1/Q3 all equal to 19 confirm that value 19 dominates non-null entries, while values 5 and 22 appear rarely enough to generate 133 outliers (14.3% of non-null rows) and strong negative skew (−2.43).

Treatment: Treat as a sparse categorical flag; convert non-null values to a one-hot or ordinal indicator and consider a separate binary 'has_footnote' feature given the 82.88% null rate.