Hathr.AI found 98.8% of NTA comorbidities in 30 scanned SNF referral packets

What did Hathr.AI find in 30 scanned SNF referral packets?

Hathr.AI listed 158 of the 160 non-therapy ancillary (NTA) comorbidities seeded across 30 scanned, faxed hospital referral packets: 98.8%, with page citations and no human review. The packets ran 60 to 241 pages, 3,820 pages in all. Hathr.AI read each one in a median of 1.7 minutes. It listed all 46 seeded primary diagnosis candidates, all 24 swallowing and diet items and all 26 special treatments.

• NTA comorbidities: 98.8% (158 of 160); per-packet median 100%
• Primary diagnosis candidates: 100% (46 of 46), including all 5 return-to-provider codes
• Swallowing and diet (SLP): 100% (24 of 24)
• Special treatments (feeding tubes, isolation, IV medication, oxygen and others): 100% (26 of 26)
• Section GG discharge function scores: 96.7% (58 of 60)
• Resolved or historical conditions kept off the active list: 73 of 74
• Processing time per packet, median: 1.7 minutes (short packets 1.1, medium 1.8, long 2.3)
• Run date: October 23, 2026

Every packet was synthetic. The patients were generated with Synthea, the open-source synthetic patient generator developed by MITRE, a federally funded research and development center, so the records carry realistic variety, formatting and clinical complexity without a single real patient.

What does 98.8% recall mean for PDPM payment?

It means Hathr.AI's list supported the right NTA payment group for every resident. The seeded comorbidities carried 225 NTA points across the 30 packets, and Hathr.AI found items worth 223 of them. Both misses were the same 1-point condition, a stage 4 sacral pressure ulcer. Both came in packets already in NTA group NC, and both stayed in NC without it: one dropped from 7 points to 6, the other from 8 to 7. On the seeded comorbidities, the NTA group Hathr.AI's list supports matched the answer key in 30 of 30 packets.

That is the number that matters to a skilled nursing facility. Under PDPM, the NTA component is set by total comorbidity points, and the step from one NTA group to the next changes the daily rate for the whole stay. A missed diagnosis that crosses a group boundary costs money every day. A missed diagnosis that doesn't still leaves the MDS incomplete. Hathr.AI's misses fell in the second category, and the cited list makes them easy for a coordinator to catch.

Hathr.AI does not code the MDS. It lists the evidence with page citations and exact quotes. A licensed coordinator verifies each item and makes every coding decision.

How does the PDPM NTA comorbidity score work?

The NTA component pays for the drugs, supplies and services a resident's conditions drive. CMS scores it from 50 conditions and extensive services, each worth 1 to 8 points, and the point total places the resident in one of six NTA groups. A diagnosis counts only when the ICD-10-CM code entered at MDS item I8000 appears on the CMS PDPM NTA mapping file.

The NTA rate also carries a variable per diem adjustment: days 1 to 3 of a Part A stay pay three times the NTA rate, and day 4 onward pays the standard rate. So a comorbidity missed in the admission packet costs the most in the first three days. At FY 2027 rates, one NTA group step is worth about $800 to $2,300 over an average 28-day stay (mean about $1,650). At $47 per user per month as of October 2026 ($564 a year, see pricing), one Hathr.AI seat breaks even on roughly one captured boundary-crossing comorbidity every three years. Those are scenario figures from CMS rates, not measured payments.

How were the test packets built?

We built the packets so that nobody, including us, could tune the result. Every packet is a realistic hospital referral with known answers hidden in it, then degraded the way a real fax degrades.

• Patients: 500 synthetic patients generated with Synthea v4.0.0. Of those, 30 were selected to cover the PDPM conditions that matter, with 30 spares.
• Answer keys first: each packet's 9 to 17 keyed items were written and committed to version control before a single page was rendered. They were then locked with SHA-256 hashes, which the scorer verified.
• Coding source: every NTA comorbidity was coded from the CMS FY 2027 PDPM ICD-10 mappings, downloaded from cms.gov.
• Realistic packets: 13 page types, including face sheets, transfer forms, H&Ps, discharge summaries, consults, progress notes, MARs, nursing flowsheets, labs, imaging, PT/OT/SLP notes, nutrition and case management. The sizes were 10 short (60 to 99 pages), 15 medium (100 to 174) and 5 long (175 to 250).
• Needles in haystacks: each keyed comorbidity is stated only on its planned page or pages, such as one problem list or one consult note. Supporting signals, such as insulin on the MAR, appear elsewhere, as they would in a real chart.
• Traps: 74 resolved or historical conditions (history of, healed, remote) were planted as distractors. A correct review leaves them off the active list.
• Fax-grade degradation: every page was printed, rescanned and faxed in simulation. That meant 200 dpi grayscale, 0.5 to 3 degree skew, noise, blur, fax headers and stamps. Each packet was rebuilt as an image-only PDF with no text layer or metadata, 4.8 to 18.7 MB.
• Validation: all 30 packets passed. Every evidence statement was confirmed OCR-legible on its keyed page, and no answer leaked through text layers or metadata. The build is deterministic: rebuilt packets are byte-identical.

How did Hathr.AI read the packets?

Hathr.AI read every page of every packet through its production HIPAA-compliant API, using one fixed prompt written before the run and never tuned on these packets. The prompt asks for what an MDS coordinator needs:

• Both primary diagnosis candidates, with return-to-provider codes flagged
• Every active NTA comorbidity on the CMS list
• Swallowing and diet items, Section GG discharge scores and special treatments
• For each item, the exact quote and every page where it is documented
• Resolved conditions in a separate do-not-code list

Rules: List only current, active items. Put resolved, healed, remote or history-of conditions in a separate short list titled Resolved or historical - do not code. For each item, quote the record's exact wording. Give an ICD-10-CM code only if the record writes one. Do not assign MDS codes, and do not add any condition the record does not state. Mark an item uncertain when the record hedges or conflicts, and say why.

The packets were 4.8 to 18.7 MB, and the API accepts one inline document of up to 4.5 MB per request. So each packet was split into parts of at most 40 pages and 4.2 MB, with pages copied byte for byte. Every part went to the model whole, with the same prompt. A final text-only request merged the part tables into one list. Because every page carries its fax header, the citations stay global across parts.

• Requests: 140 (110 part requests and 30 merges), 0 errors and 0 rate-limit events
• Part request, median: 19.9 seconds (p95 23.8 s, max 28.0 s)
• Merge request, median: 26.9 seconds (max 34.5 s)
• Whole run: 30 packets, one at a time, in 50 minutes
• Settings: API default temperature (0.2), one run per packet

Why did reading every page beat search?

Reading every page found 98.8% of NTA comorbidities, against 39.4% when the same packets went through upload and search. We ran both on October 23, 2026, with the same 30 packets, the same prompt and the same scorer, and the gap is the most useful thing this benchmark taught us.

Search-based retrieval sends the model only the passages that match the query. A primary diagnosis is repeated on the face sheet, the H&P and the discharge summary, so search finds it. A comorbidity is often stated once, on one problem list or one consult note deep in the packet, and search never sends that page to the model. Whole-packet reading costs some precision: the list carries more rows for the coordinator to strike. For PDPM work the trade favors recall, because striking a row takes seconds on the cited page, while a missed comorbidity can drop the NTA group.

Which items did Hathr.AI miss?

Hathr.AI missed 4 of the 316 items it was expected to list (98.7% recall), and put 1 of 74 planted resolved conditions on the active list.

• Stage 4 sacral pressure ulcer, twice: one in a 95-page packet, one in a 184-page packet. Each was worth 1 NTA point, and neither changed the NTA group.
• Section GG discharge scores, mobility and self-care (once each): both in the same 208-page packet.
• One resolved condition listed as active, in a 95-page packet. It was the only distractor out of 74 that reached the active list.

We report every miss against the answer key as written, with no exclusions.

What does this mean for an MDS coordinator or admissions team?

It means the first pass through a packet no longer has to be a page-by-page read. Hathr.AI produces a cited list in about two minutes, and the MDS coordinator's job becomes verification: open each cited page, confirm or strike the row, and code. Function items still follow the Section GG scoring rules.

That matches what admissions teams tell us. One central admissions team reviews referrals for 12 skilled nursing facilities. When hospital auto-referral systems started sending overnight batches, each reviewer's load rose from about 12 packets a day to about 40. The details they most feared missing were feeding tube formulas, isolation needs and expensive medications. In this benchmark, Hathr.AI listed all 26 seeded special treatments across the 30 packets.

Hathr.AI customers report saving 45 to 90 minutes of coordinator time per Medicare Part A admission. That figure comes from customers; this benchmark measures something different, Hathr.AI's own processing time of 1.7 minutes per packet. A timed comparison against a manual review is the next stage of this study, and we will publish it here.

See how skilled nursing teams use Hathr.AI, or upload a referral packet and get every PDPM item with its page citation, free for 7 days.

How is Hathr.AI different from other SNF admissions tools?

Hathr.AI publishes what it finds, not just how fast it reads. A leading SNF admissions platform we reviewed advertises packet review in under a minute and document counts in the millions, but publishes no recall figure, no method and no list of misses. Speed is easy to claim; recall on a needle-in-a-haystack comorbidity is what changes the payment.

• Measured recall with a published method: 98.8% NTA recall on 30 locked-key packets, with every miss listed.
• Whole-packet reading: every page reaches the model, up to the 241-page packets in this test. The Hathr.AI app holds 90+ files and 100,000+ pages in one workflow.
• Built for scanned faxes: image-only, skewed, fax-degraded PDFs with no text layer.
• Evidence, not codes: exact quotes and page citations, with the coordinator making every coding decision.
• Compliance first: Hathr.AI runs in AWS GovCloud, hosted on FedRAMP High infrastructure. A signed BAA is included on every HIPAA-compliant plan, and customer data never trains a model.
• Simple pricing: the Hathr.AI app is $47 per user per month as of October 2026, with no seat minimums and a 7-day free trial (pricing).

Hathr.AI continually improves the models and pipeline behind its platform. This page reports the results measured on October 23, 2026, and we will publish new runs here, using the same locked packets and scorer.

Frequently asked questions

Can AI find NTA comorbidities in a hospital referral packet?

Yes, when it reads the whole packet. In a 30-packet synthetic benchmark, Hathr.AI listed 98.8% of seeded NTA comorbidities (158 of 160) from scanned, faxed packets of 60 to 241 pages, with page citations, in a median 1.7 minutes per packet.

What are NTA comorbidities under PDPM?

NTA comorbidities are the 50 conditions and extensive services CMS uses to score the non-therapy ancillary component of PDPM. Each is worth 1 to 8 points, and the total sets the resident's NTA group. A diagnosis counts only if its ICD-10-CM code is on the CMS PDPM NTA mapping file.

What are the PDPM NTA groups?

There are six: NA (12 or more points), NB (9 to 11), NC (6 to 8), ND (3 to 5), NE (1 to 2) and NF (0). Days 1 to 3 of a Part A stay pay three times the NTA rate, so a missed comorbidity that crosses a group boundary costs the most at admission.

Did Hathr.AI's misses change the PDPM NTA group?

No. Both missed comorbidities were 1-point stage 4 sacral pressure ulcers in packets already in NTA group NC, and both stayed in NC without them. On the seeded comorbidities, the NTA group Hathr.AI's list supports matched the answer key in all 30 packets.

How long does Hathr.AI take to review a referral packet?

Hathr.AI processed each packet in a median 1.7 minutes: about 1.1 minutes for 60 to 99 pages, 1.8 for 100 to 174, and 2.3 for 175 to 250. That is processing time, not coordinator time. Hathr.AI customers report saving 45 to 90 minutes of coordinator time per Medicare Part A admission.

Does Hathr.AI code the MDS?

No. Hathr.AI lists each PDPM-relevant item with the record's exact wording and the pages where it appears. The MDS coordinator verifies the evidence and makes every coding decision.

Was real patient data used?

No. Every packet was built from Synthea synthetic patients and stamped SYNTHETIC TEST RECORD on every page. For real referrals, Hathr.AI runs in AWS GovCloud, hosted on FedRAMP High infrastructure, with a signed BAA on every HIPAA-compliant plan.

Why does reading the whole packet matter?

Comorbidities are often stated once, on a single page deep in the packet. Search-based retrieval sends the model only the passages that match a query and misses many of them: 39.4% NTA recall in our test, against 98.8% when every page reached the model.

Related reading

• What is PDPM? The six components of the SNF Part A rate
• Section GG scoring guide
• MDS coordinator guide
• AI in nursing homes
• AI medical record review

Sources and method

• CMS. FY 2027 PDPM ICD-10 mappings and the PDPM NTA Comorbidity Score fact sheet, downloaded from cms.gov on October 15, 2026.
• CMS. Long-Term Care Facility Resident Assessment Instrument User's Manual, Chapter 6 (PDPM classification).
• LeadingAge. SNF final rule overview: CMS finalizes 2.4% payment update for FY 2027.
• Synthea. Synthetic patient generator, v4.0.0, developed by MITRE.

Method note: all Hathr.AI figures come from two runs on the production Hathr.AI API on October 23, 2026: whole-packet (18:15 to 19:05 UTC) and upload with retrieval. Both used the same 30 locked-key packets, the same frozen prompt and the same automated scorer (rules frozen at commit f11aa3f).