Analyze patient data in ChatGPT without compromising PHI

With Subsalt’s ChatGPT plugin, everyone at your hospital can get patient population insights as fast as they can ask a question. That’s the power of curiosity, de-identification, and AI.

You deployed ChatGPT across your health system. Now you have to prove its value.

Data and AI leaders work hard to deploy AI tools quickly and safely. Now finance is scrutinizing AI spend, and the next purchase depends on ROI.

But ChatGPT isn’t connected to patient data, so the value of one of your biggest AI implementations is difficult to quantify. Your operations team still depends on the overextended analytics team for answers. Shadow AI is still a problem. And the potential of AI is limited to people with PHI access, because patient privacy is paramount.

Healthcare deserves a scalable, simple way to help humans and AI learn faster from the patient data it produces every day, without exposing PHI. That belief is why Subsalt exists. It’s why Subsalt engaged Datavant to independently evaluate its de-identification technology under HIPAA’s Expert Determination standard.

And it’s why Subsalt built its plugin for ChatGPT.

The Reality Data Leaders Are Left to Manage The Reality Data Leaders Can Deliver

  • Healthcare AI spend under scrutiny Leaders who can point to what their AI investment unlocked
  • AI promises and potential outpacing what the data allows A learning health system in practice, not just strategic vision
  • Governance teams forced to manage multiple specialized AI tools Data leaders who can approve new uses of patient data without taking on new PHI risk
  • Every new AI use of patient data reviewed case-by-case Research and quality teams acting on insight while it’s still timely
  • Employees putting PHI in unauthorized AI tools Population-level insight open to everyone, not only the PHI-cleared few
  • Patient data questions still routed to an overextended analytics team An analytics team freed for its most strategic work
Show the reality data leaders face
Choose a use case

See the Subsalt plugin for ChatGPT in action

Insight shouldn’t wait in a queue. With the Subsalt plugin, any team can explore questions in ChatGPT the moment they arise, so data leaders can scale insight across the organization without adding PHI risk.

Pick a clip to watch the Subsalt plugin in action.

Illustrative example. Everything in the window is simulated, including the data, names, numbers and source labels.

Simulated conversation shown in the video. All data, names and numbers are illustrative.

Question: @Subsalt How have avoidable days trended over the last 12 months for patients who were medically ready for discharge? What's driving the excess LOS, and where should we focus to bring it down?

Subsalt:

Avoidable days held near 460 a month through July, then climbed to 840 in October. This quarter's total, 2,240 days, is up 54% from last quarter.

Avoidable days by month (days patients waited after the medically ready flag, last 12 months): Nov 430, Dec 450, Jan 480, Feb 460, Mar 450, Apr 455, May 470, Jun 480, Jul 500, Aug 640, Sep 760, Oct 840.

Most of the increase comes from patients waiting on post-acute care:

  • Waiting for a SNF bed: +480 days, more than half the increase
  • Home-health setup: +170 days
  • Inpatient rehab acceptance: +80 days

Where to focus:

  • Start SNF referrals at admission for patients likely to need one
  • Escalate the same day when a medically ready patient has no accepted bed
  • Book home-health intake before the discharge order is written

If you want, I can break this down by nursing unit or payer.

Follow-up: Break that down by nursing unit. Which units carry the most avoidable days, and what's the main barrier on each?

Subsalt:

5 West Medicine carries about a third of all avoidable days. Most are patients 75 and older waiting for a SNF bed.

Avoidable days by nursing unit, this quarter:

  • 5 West Medicine: 720 (SNF bed waits)
  • 4 East Cardiac: 410 (Home-health setup)
  • 6 North Ortho: 360 (Rehab acceptance)
  • 3 South Surgical: 300 (Payer authorization)
  • 7 West Oncology: 250 (Hospice placement)
  • All other units: 200 (Mixed)

5 West is the best place to pilot earlier SNF referrals. On 4 East, booking home-health intake sooner would help most.

Illustrative example. Everything in the window is simulated, including the data, names, numbers and source labels.

Simulated conversation shown in the video. All data, names and numbers are illustrative.

Question: @Subsalt How has our initial denial rate trended over the last 12 months? Which denial reasons are driving the increase, and what should we fix first?

Subsalt:

Your initial denial rate held near 8.4% through April, then climbed to 11.2% in October. Over the last six months, payers denied 2,986 claims.

Initial denial rate by month (share of claims denied on first submission, last 12 months): Nov 8.3%, Dec 8.5%, Jan 8.4%, Feb 8.3%, Mar 8.4%, Apr 8.4%, May 9.1%, Jun 9.6%, Jul 10.0%, Aug 10.4%, Sep 10.8%, Oct 11.2%.

Three reasons account for most of those denials:

  • Missing prior authorization: 1,193 claims, 51% from Larkspur Health Plan
  • Missing or invalid patient information: 859 claims
  • Medical necessity: 762 claims

What to fix first:

  • Check authorization at scheduling, starting with outpatient imaging
  • Add each payer's rules to your claim scrubber
  • Use a medical-necessity checklist for high-denial services

If you want, I can break this down by payer.

Follow-up: Break that down by payer. Which payers deny the most, and for what reason?

Subsalt:

Larkspur Health Plan accounts for 36% of denials, mostly for missing prior authorization. Crestview Medicare Advantage denies most often for medical necessity.

Denied claims by payer, last six months:

  • Larkspur Health Plan: 1,078 (Missing prior authorization)
  • Crestview Medicare Advantage: 785 (Medical necessity)
  • Bayfield Community Health Plan: 600 (Missing or invalid info)
  • All other payers: 523 (Mixed)

Start with imaging orders for Larkspur Health Plan, before the patient arrives. For Crestview, a medical-necessity checklist at order entry would help most.

Illustrative example. Everything in the window is simulated, including the data, names, numbers and source labels.

Simulated conversation shown in the video. All data, names and numbers are illustrative.

Question: @Subsalt Sepsis is coming into HRRP. How have our 30-day readmissions after sepsis trended over the last 12 months, Medicare Advantage included? What's driving them, and where should we focus first?

Subsalt:

Your 30-day readmission rate after sepsis rose from 17.9% to 21.4% over the last 12 months. This quarter, Medicare Advantage patients readmitted at 23.6%, versus 19.8% in traditional Medicare.

30-day readmissions after sepsis, by month (all-cause rate, Medicare Advantage included, last 12 months): Nov 17.9%, Dec 18.2%, Jan 18.0%, Feb 18.6%, Mar 18.4%, Apr 18.9%, May 19.1%, Jun 19.6%, Jul 19.8%, Aug 20.5%, Sep 21.0%, Oct 21.4%.

Most readmissions trace to three causes:

  • Recurrent or new infection: 38% of readmissions, most within 7 days
  • Heart failure or fluid overload: 17%
  • Acute kidney injury: 11%

Where to focus:

  • Book a follow-up visit within 7 days of every sepsis discharge
  • Send the antibiotic plan and repeat-lab orders with every SNF transfer
  • Recheck weight and kidney function within 72 hours of discharge

If you want, I can break this down by discharge destination.

Follow-up: Break that down by discharge destination. Where do readmissions concentrate, and what's the top cause for each?

Subsalt:

Patients discharged to a skilled nursing facility readmit at 27.8%, the highest rate, and infection is the top cause. Patients who go home without services readmit least, at 14.2%.

Readmission rate by discharge destination, this quarter:

  • Skilled nursing facility: 27.8% (Recurrent infection)
  • Home with home health: 21.5% (Heart failure)
  • Inpatient rehab: 16.9% (Acute kidney injury)
  • Home, self-care: 14.2% (Recurrent infection)

Start with your SNF partners. A shared sepsis handoff and a 72-hour check-in would reach the patients most likely to come back.

Illustrative example. Everything in the window is simulated, including the data, names, numbers and source labels.

Simulated conversation shown in the video. All data, names and numbers are illustrative.

Question: @Subsalt We're weighing a Phase 3 MASH trial. How many patients could prescreen in from the record: BMI 27 or higher, FIB-4 of 1.3 or above, no cirrhosis and no alcohol use disorder? Which criteria cut the pool most?

Subsalt:

7,140 patients could prescreen in from the record. The biggest cut isn't fibrosis risk. It's missing labs: only 39% of patients with a BMI of 27 or higher had the liver panel and platelet count needed to calculate FIB-4.

Patients who could prescreen in, by criterion:

  • BMI 27 or higher: 142,600
  • FIB-4 can be calculated: 55,600 (39% have the labs)
  • FIB-4 of 1.3 or higher: 9,850
  • No cirrhosis: 8,320
  • No alcohol use disorder: 7,140 (Prescreen pool)

Where the pool shrinks:

  • FIB-4 can't be calculated: 87,000 patients lack a liver panel or platelets
  • FIB-4 below 1.3: 45,750 more
  • Cirrhosis or alcohol use disorder: 2,710 more

Where to focus:

  • Add FIB-4 to primary care lab panels for patients with obesity or diabetes
  • Send patients with FIB-4 of 1.3 or higher for elastography before screening
  • Budget for screen failure, since the record can't stage fibrosis

If you want, I can break this down by site.

Follow-up: Break that down by site. Which sites have the largest pools, and what limits each one?

Subsalt:

Main campus holds 44% of the prescreen pool. At the community sites, missing labs are the biggest limit.

Prescreen pool by site:

  • Main campus: 3,140 (FIB-4 below 1.3)
  • Harbor Street Clinic: 1,580 (Missing labs)
  • Lakeview Family Medicine: 1,320 (Missing labs)
  • Westgate Clinic: 1,100 (Missing labs)

Start site activation at main campus and Harbor Street. A FIB-4 lab order set at Lakeview and Westgate would grow the pool fastest.

ChatGPT and PHI How Subsalt protects patient data

Subsalt operates securely inside the health system’s environment, analyzing patient data through secure compute technology that isolates PHI in a “black box.” ChatGPT receives only de-identified, aggregate results that pass Subsalt’s five privacy checks, ensuring each answer preserves patient privacy.

Request a Demo
How a question moves: a staff member asks in ChatGPT, which is hosted by OpenAI. The question reaches Subsalt Secure Runtime inside the health system’s environment through the Subsalt plugin. Secure Runtime reads patient records from the health data warehouse and analyzes them inside the health system, then checks the output against five privacy checks. The patient records stay inside the health system. Only a de-identified, aggregate result returns to ChatGPT.

FAQs

Data governance and compliance leaders tend to raise the same questions at the outset. The answers below cover the analysis the Subsalt plugin supports, where it runs and how Subsalt safeguards PHI.

Learn more about Subsalt for Health Systems

Can ChatGPT analyze hospital operations data?

Yes. With the Subsalt plugin, healthcare leaders can query operational data in ChatGPT, within the scope of the health system’s configured data and analysis.

Can hospitals analyze claim denials in ChatGPT?

Yes. The Subsalt plugin supports analysis of claim-denial patterns, including comparisons by payer, service line, and denial reason, wherever the relevant tables and fields are available to Subsalt.

How does the Subsalt ChatGPT plugin analyze health system data?

When a question requires analysis of PHI, Subsalt translates the prompt into SQL and executes it in Secure Runtime, so PHI is never exposed. ChatGPT receives only de-identified, aggregate results that pass Subsalt’s privacy checks.

Does ChatGPT have access to patient data or PHI?

No. Subsalt shares only de-identified answers with ChatGPT, supporting health systems’ compliance with the HIPAA Privacy Rule. With the Subsalt plugin, ChatGPT never has access to PHI.

How does Subsalt protect PHI?

Subsalt operates entirely within a health system’s secure environment. The Subsalt plugin connects ChatGPT to Subsalt Secure Runtime, an isolated “black box” that keeps PHI hidden from end users. Secure Runtime translates each question into SQL, executes it against PHI, and returns accurate, de-identified, aggregate results. ChatGPT receives only results that pass each of Subsalt’s five privacy checks.

Has Subsalt’s de-identification been independently evaluated?

Yes. Datavant independently evaluated Subsalt’s de-identification technology under HIPAA’s Expert Determination standard.

Read what Datavant evaluated

What questions can healthcare finance teams explore with ChatGPT?

Finance teams can examine claim-denial patterns by payer, service line, and reason, then refine the analysis through follow-up questions. Available results depend on the health system’s tables, fields, and configuration.

See Subsalt work with ChatGPT live: Schedule a Demo Today

Population-level insight, open to everyone across the health system, with PHI exactly where it belongs. Request a demo to see it in ChatGPT.

Request a Demo
  1. Request a DemoSchedule a call with the Subsalt team.
  2. Bring a Question Stuck in the QueuePick an analysis that matters to the health system but keeps waiting on data access.
  3. Explore the Possibilities with SubsaltSee how Subsalt would answer it in ChatGPT, with PHI kept inside the health system’s environment.
Get the Subsalt plugin (opens in a new tab)

A Subsalt team member will follow up to coordinate a time.

Request a Demo

Tell us what you’re working to advance. One member of our team will follow up by email to schedule a time.

By submitting, you agree that Subsalt may store and process this information to respond to your request. See our privacy policy.

Request received

We’ll be in touch.

A member of our team will email you to coordinate a time.