AI Trust OSBlogResearch
Research2026-04-29· 7 min read

Zero-Egress Psychiatric AI: On-Device LLM Deployment for Privacy-Preserving Mental Health Support

Published on arXiv. A privacy-first AI platform for mental health decision support — all patient data processed entirely on-device. No data ever leaves the device, no cloud inference, no privacy risk.

📄 Published Research — Peer-reviewed paper published on arXiv.
Full paper: arxiv.org/abs/2604.18302

Privacy is one of the most critical yet underaddressed barriers to AI adoption in mental healthcare — particularly in high-sensitivity environments such as military, correctional, and remote healthcare settings, where the risk of patient data exposure can deter help-seeking behaviour entirely.

Existing AI psychiatric decision support systems predominantly rely on cloud-based inference pipelines, requiring sensitive patient data to leave the device and traverse external servers. In contexts where a patient's willingness to seek help depends on absolute confidence that their data never leaves their hands, this architecture is unacceptable.

The Zero-Egress Principle

The zero-egress architecture guarantees a single property: no patient data is transmitted to, processed by, or stored on any external server at any stage.

This is not a privacy policy. It is an architectural guarantee — enforced by the absence of any external network call during inference. The platform runs entirely on the device. There is nothing to breach externally because nothing leaves.

The On-Device Model Consortium

The platform integrates three lightweight, fine-tuned, quantized open-source LLMs working in ensemble:

  • Gemma — compact architecture, strong instruction following
  • Phi-3.5-mini — optimised for resource-constrained mobile hardware
  • Qwen2 — multilingual, for cross-language clinical applicability

An on-device orchestration layer coordinates ensemble inference with consensus-based diagnostic reasoning, producing DSM-5-aligned assessments without any cloud dependency.

Clinical Capabilities

  • Clinician decision support — differential diagnosis assistance and evidence-linked symptom mapping against DSM-5 criteria
  • Patient-facing self-screening — accessible mental health screening with clinical safeguards, for environments where clinician access is limited or delayed

Performance on Commodity Hardware

The evaluation demonstrates diagnostic accuracy comparable to server-side predecessors while sustaining real-time inference latency on commodity mobile hardware — the class of device available in correctional facilities, remote postings, and field environments.

Regulatory Significance

  • HIPAA — Protected Health Information never enters a covered cloud environment. Business Associate Agreements become structurally unnecessary.
  • GDPR — No personal data leaves the jurisdiction of the patient's device. Cross-border transfer restrictions do not apply.
  • EU AI Act — On-device medical AI may qualify for reduced regulatory burden given the absence of centralised data processing.

The Broader Implication

This research demonstrates that the assumed trade-off between AI capability and privacy is not architectural — it is a design choice. For sensitive domains where privacy is non-negotiable, on-device AI with quantized open-source models now provides capability parity without the privacy risk. The zero-egress principle has direct applicability beyond healthcare: legal, financial, defence, and government AI deployments all face the same data sovereignty requirements.

Read the full paper: arxiv.org/abs/2604.18302

PrivacyOn-Device AIHealthcare AIHIPAAZero TrustLLMResearch

Automate ISO 42001 and EU AI Act compliance

AI Trust OS maps your AI systems to every framework automatically. Free to start, no credit card required.

Get Started FreeBack to Blog

Related Articles

What is ISO 42001? The AI Management System Standard Explained

7 min read

EU AI Act Compliance Guide 2025: What Your Company Needs to Know

9 min read

What is an AI Registry? And Why the EU AI Act Requires One

7 min read