Private LLM for healthcare: your data, your environment
A model fine-tuned on your own records and run in your cloud or data centre, with the use cases and data handling explained.
Where a private model fits (and where it doesn't)
| Use case | Input | Output | Approach | Human review |
|---|---|---|---|---|
| Prior-authorization packet assembly | Clinical notes, payer criteria | Draft request mapped to criteria | Fine-tune for format; retrieve current payer policy | Always — staff submit |
| Discharge summary first draft | Inpatient notes | Draft in your template | Fine-tune | Always — clinician signs |
| Claims denial triage | Denial codes, claim notes | Category, next action | Fine-tune | Sample QA |
| Patient-message routing | Portal messages | Route + urgency flag | Fine-tune | Urgent flags reviewed |
| Policy and formulary Q&A | Internal policies, formularies | Answer with citation | Retrieval (facts change) | Spot checks |
A worked example: claims denial triage
A revenue-cycle team already sorts denials into a few dozen categories and decides what to do next. We take past denials with the category and action your staff chose, remove identifiers that the task doesn't need, and fine-tune a model in your environment to make the same calls in the same format. We test it on recent denials it hasn't seen, against your current process, with a pass mark agreed in week 1. It then runs alongside the team with a weekly review sample. Nothing leaves your network, and the model is versioned so you can roll it back.
See how this would work on your data.
Book a callData you'd use, and how it's handled
| Data | Sensitivity | Handling |
|---|---|---|
| Clinical notes, summaries | PHI (US) / health data (EU special category) | De-identify for training where the task allows; stays in your environment on customer-VPC deployments |
| Claims and billing data | PHI | Same |
| Policies, payer criteria, formularies | Internal | Retrieval, versioned |
| Staff decisions (labels) | Internal | Training labels after review |
Rules depend on where you operate: in the US, HIPAA (see our HIPAA page); in the EU, health data is a special category under GDPR Article 9 (see Private LLM for GDPR); in India, the DPDP Act.
Where it runs
Your cloud account (fastest start), or on-prem as a scoped data-centre install in a managed engagement. See on-prem and VPC deployment.
What we'd do first
- 1Pick one administrative workflow with a clear right answer and plenty of past examples — denial triage or message routing are good candidates.
- 2Agree the pass mark, build the test set from your own cases, and reach a canary release in 8 weeks.
- 3Clinical-documentation drafting comes next, once the review process is proven.
- 4See the 8-week pilot.
Frequently asked questions
Can a private LLM be used for clinical decisions?
We don't start there. Software intended to diagnose or treat may be a medical device under FDA rules (FDA); first projects should be administrative with human sign-off.
Does patient data leave our network?
It depends on the deployment. Customer-VPC deployments keep training and inference in your environment, and on-prem runs as a scoped data-centre install in a managed engagement. IRIS8-hosted runs the app on GCP and inference on Vultr.
Do we need lots of data?
For one narrow task, past examples your staff already produced are usually the starting point. We check volume and quality in week 1.
What does HIPAA mean for it?
HIPAA applies to how your organization uses it; see what HIPAA requires of an LLM.
Who owns the model?
You do.