Patient Mix, decoded

Where the data comes from

The view is built on the CMS Minimum Data Set (MDS) Facility Frequency Report, a publicly published, quarterly dataset aggregating every resident assessment CMS receives — about 498 MDS items per facility per quarter, broken out by Overall % / Short-Stay % / Long-Stay %. Crucially, this is all-payer: it covers Medicare Advantage, traditional Medicare, Medicaid, and private pay alike. That's a sharper lens for behavioral and memory-care identification than Medicare-only sources, where many of those residents are excluded by definition.

Source dataset: CMS MDS Facility Frequency. SNF Shark ingests the latest published quarter, filters down to ~27 curated items (see below), pivots into one row per facility per quarter, and pre-computes a national benchmark for every column. The card you see on Operating Metrics is reading that one pivoted row.

What we surface (the curated 27)

MDS publishes far more than a financial-diligence persona needs. We curated 27 items across four buckets — chosen because each one either (a) helps classify the facility, or (b) materially changes the underwriting question the user is asking. Items pulled directly via their MDS code prefix:

Age distribution (Section A)

MDS publishes 6 buckets (0–30, 31–64, 65–74, 75–84, 85–95, 96–999); we collapse to four. The under-65 column carries most of the diagnostic weight — paired with elevated schizophrenia coding it's the strongest signal we have that this is a genuinely younger / behavioral facility rather than an over-coded general SNF.

• A0900 Age of resident — CMS publishes six cells (0–30, 31–64, 65–74, 75–84, 85–95, 96+), which we collapse to 4 buckets: Under 65, 65–74, 75–84, 85+

How we surface it. The Patient Mix card leads with the under-65 share compared to the facility's state peer (the same benchmark used elsewhere on the card). We raise a younger / behavioral flag only when the under-65 share is reported, at least 25%, and at least 2× the peer — a deliberately high bar that fires for roughly the top fifth of facilities, not every facility running slightly above its state. The flag is a demographic prompt to read the diagnosis and behavior rates below, not a clinical determination.

Suppression applies per source cell: CMS withholds any of the six age cells that holds fewer than 11 residents. For roughly two-thirds of facilities both under-65 cells (0–30 and 31–64) are withheld; we show that as "too few residents under 65 to report" rather than a blank. The combined under-65 count can still reach 20 (two cells of up to 10 each), so we don't claim it is under 11 — but for a general SNF the suppression usually reads as the answer: a predominantly 65+ census.

Cognitive & behavioral diagnoses (Section I, "Yes")

• I4200 Alzheimer's
• I4800 Non-Alzheimer's dementia
• I6000 Schizophrenia (coding caveat — see below)
• I5800 Depression (other than bipolar)
• I5900 Bipolar disorder
• I5700 Anxiety disorder
• I5950 Psychotic disorder (other than schizophrenia)
• I6100 PTSD

Physical comorbidities (Section I, "Yes")

• I0600 Heart Failure (CHF) · I6200 COPD · I2900 Diabetes
• I4500 Stroke / CVA · I5300 Parkinson's · I1500 Renal failure / ESRD

Behavioral indicators (Section E, "presence")

We invert the "Behavior not exhibited" rate to get presence. So "Physical aggression: 12%" means 12% of residents exhibited physical aggression at any frequency during the assessment.

• E0200A Physical aggression · E0200B Verbal aggression
• E0800 Rejection of care · E0900 Wandering
• E0100A Hallucinations · E0100B Delusions

Acuity profile (Section O, "While a Resident")

• O0110K1B Hospice · O0110F1B Mechanical ventilator (invasive)
• O0110J1B Dialysis · O0110H1B IV medications
• O0110E1B Tracheostomy · O0110C1B Oxygen therapy · O0110M1B Isolation / quarantine

The four classifications + thresholds

We compute a single facility-level classification at backfill time and cache it on the row. The pill on the Snapshot tab uses this directly. Thresholds are intentionally conservative — we'd rather under-classify a typical SNF than mislabel one that's slightly above national average.

Precedence when multiple triggers fire: memory-care > behavioral-heavy > high-acuity.

The clinical fingerprint radar

The radar on the Patient Mix card plots raw prevalence percentages — the same numbers as the deviation bars beside it, in a second geometry. Nothing is normalized, standardized, or re-scaled per axis: we deliberately rejected z-scores (the benchmark stores means only, and the suppression-censored, zero-inflated distributions would make a "standard deviations from state mean" radar encode facility size rather than clinical mix) and ratio scaling (state means for rare diagnoses are pulled toward zero by the suppressed-as-0 convention, so ratios explode exactly where the data is weakest).

Fixed axes, fixed order

Every facility gets the same 8 axes in the same positions, so shapes are comparable across facility pages. The default fingerprint lens runs clockwise from the top: depression, anxiety, non-Alzheimer's dementia (the mood/cognitive arc), then diabetes, heart failure, COPD, renal failure/ESRD, and stroke (the chronic-disease arc). These 8 were chosen for reporting coverage: they are the most consistently published items, so the shape reflects patient mix, not CMS suppression patterns. Note that Alzheimer's is not on the fingerprint (it is suppressed at ~3 in 4 facilities) — the non-Alzheimer's axis alone is not a total-dementia read; use the Behavioral lens or the All-metrics list for that.

When it renders, and the Behavioral lens

The fingerprint renders only when at least 6 of its 8 axes are CMS-reported for the facility (about 72% of facilities with a published MDS profile clear that floor at the Q1 2026 vintage). The Behavioral lens has a deliberately lower floor: at least 5 of its 8 items reported, of which at least one is a serious-mental-illness item (schizophrenia, bipolar, or other psychotic disorder). The lens's 8 items are Alzheimer's, non-Alzheimer's dementia, depression, anxiety, schizophrenia, bipolar, other psychotic disorder, and PTSD — and its target population, behavioral-heavy facilities, systematically has the rarest of those (Alzheimer's, other psychotic, PTSD) suppressed. At a 6-of-8 floor the lens unlocked for only 24% of behavioral-heavy facilities — the audience it exists for; the 5-of-8 + SMI-reported rule unlocks it for 44% of them (about 15% of all profiled facilities), counting only facilities where the radar renders at all. The fingerprint floor governs rendering: a facility that clears the behavioral floor but not the fingerprint floor shows no radar.

One caveat that comes with the lower floor: at 5 of 8, the missing spokes are almost always the rare arcs, so two facilities' behavioral shapes can be drawn over slightly different effective vocabularies. Suppressed spokes are always explicitly marked (see below), and when two or more adjacent spokes are suppressed the facility polygon's fill is dropped entirely — outline only — so no area is asserted across the gap.

Suppression, scale, and the red markers

CMS-suppressed axes are never plotted as zero and never dropped from the geometry: the spoke stays, the label shows a hatched suppressed tag (each radar axis is a single CMS cell, so the <11-residents rule applies literally), a hollow marker sits at the benchmark's value on that spoke, and the facility outline crosses the sector as a dotted bridge that asserts no value. The axis scale is shared across all 8 spokes — the largest plotted value rounded up to the next 10 — and printed on the top spoke; it is computed independently of the deviation bars' scale, so the two can differ side by side (each prints its own). A red vertex ring marks an axis only when the facility is at least 2× the benchmark and at least 5% of residents (the same rule as the red bars), and the ×-multiplier chip additionally requires the benchmark itself to be ≥2% so a near-zero denominator never prints a meaningless multiple. The benchmark polygon is the equal-weight state mean described below (national fallback for states with fewer than 5 reporters) — closed on all axes because suppressed cells count as 0 in that average.

Known caveats & what we excluded

How the "vs X% [STATE]" benchmark is computed

Bars compare the facility's value to its state average rather than national, since patient mix varies more by state than by region (Medicaid policy, IMD rules, and local demographics all matter). National is used as a fallback for very small states where the state row has fewer than 5 reporting facilities.

Facility and benchmark are compared at the same quarter: the benchmark is resolved at the facility's most recent reporting quarter, not at the newest benchmark vintage. A facility that stops appearing in newer CMS releases is therefore compared against the average from the quarter it last reported. In the rare case where no benchmark exists at that exact quarter, the nearest earlier quarter is used and the card discloses the benchmark's quarter next to the scope label.

The average treats CMS-suppressed cells (small cell counts, shown as _ in the source data) as 0, not as missing. Without this adjustment, the "average" for low-incidence items like schizophrenia or mechanical ventilator would silently mean "average among facilities that report a value", which is selection-biased upward — those facilities are exactly the ones with elevated rates. With the adjustment, the benchmark approximates the true population mean: typical facilities (with 0 or suppressed counts) pull the average toward zero, and outliers stand out clearly.

Schizophrenia coding (CMS QSO-23-05-NH)

CMS has flagged schizophrenia as a known noisy MDS item — historically over-coded by some facilities, sometimes to justify antipsychotic use. The 2023 QSO memo (QSO-23-05-NH) introduced stronger audits. We surface the % as-published and flag the bar with a tooltip, but treat very high outliers (e.g. 25%+) as worth investigating rather than taking at face value.

Cell suppression

CMS replaces values with _ when a cell would identify too few residents (small-cell suppression). We treat these as NULL— the bar and the % both render as "—". Facilities where every curated item is suppressed get no MdsFacilityProfile row at all, and the card simply auto-hides. Most of these are very small facilities or new admissions.

Age buckets add one wrinkle: our four buckets collapse six CMS cells, and suppression happens before the collapse. When one of a bucket's two source cells is suppressed (say 96+, but not 85–95), the bucket shows the reported cell only — a floor, not the full bucket. The composition bar pools all suppressed mass into a single hatched segment, which is why the hatch can hold residents belonging to any bucket, including 85+.

Per-item denominators

MDS Frequency percentages are per item: each item is computed over the residents assessed for that item, so denominators differ across items within the same facility-quarter — the age (A0900) denominator, the census-derived resident total shown on the card, and a Section I denominator can all be different numbers. This is also why long-stay + short-stay resident counts don't sum exactly to the total. Never convert a percentage back to a head-count using the displayed resident total.

Race / ethnicity

MDS publishes a 15-category resident race / ethnicity breakdown. We deliberately don't surface it — the financial-diligence persona isn't asking that question, and adding the columns would invite editorial decisions we'd rather not litigate. If you have a legitimate need for those fields, get in touch and we'll add them with the right framing.

Medications (Section N)

MDS Section N covers antipsychotic / antianxiety / antidepressant / opioid use, and CMS already surfaces antipsychotic % as a star-rating input — see QM 419 on the Clinical Dashboard. To avoid duplication, the Patient Mix card intentionally doesn't double-up on medication frequencies.

Cadence & freshness

CMS publishes MDS Facility Frequency quarterly, with roughly a one-quarter lag. SNF Shark ingests each new release as it drops — see the live freshness tracker on Data sources for the exact "our last refresh" date.