ExposureGuard-SynthRewrite-T5

Redacting PHI preserves privacy but destroys the clinical note. A note full of [REDACTED] tokens is useless for downstream processing, audit review, or secondary analysis. This model takes a different approach: it rewrites the note into safe, readable text where real PHI is replaced with consistent pseudonyms rather than blanks.

The version suffix is the key behavior. When cross-modal linkage triggers a pseudonym rotation, the version counter increments. Any system trying to join records across version boundaries hits a deliberate discontinuity. Prior and future records cannot be linked through the pseudonym chain.

---

What it does

Input: a clinical note plus exposure context (risk score, modality, pseudonym version, trigger reason). Output: a fully rewritten note with real PHI replaced by version-stamped pseudonyms.

Input:  rewrite: [risk=0.72] [modality=asr] [version=1] [trigger=cross_modal_linkage]
        note: James Smith is a 67-year-old patient presenting with chest pain.
              DOB 03/22/1955. Contact: 555-0142. MRN: MRN482910.

Output: Robert Taylor is a 67-year-old patient presenting with chest pain.
        DOB [DOB-V1]. Contact: [PHONE-V1]. MRN: [MRN-V1].

Version 0 means no trigger has fired. Version 1 means pseudonyms rotated once. The version suffix in every placeholder makes the rotation point explicit and auditable.

---

Usage

from inference import load, rewrite

model, tok = load(".")

result = rewrite(
    note="James Smith, MRN 482910, admitted with chest pain. DOB 03/22/1955.",
    risk=0.72,
    modality="asr",
    version=1,
    trigger_reason="cross_modal_linkage",
    model=model,
    tokenizer=tok,
)
print(result)

Or directly with transformers:

from transformers import T5ForConditionalGeneration, T5Tokenizer
import torch

tok   = T5Tokenizer.from_pretrained("vkatg/exposureguard-synthrewrite-t5")
model = T5ForConditionalGeneration.from_pretrained("vkatg/exposureguard-synthrewrite-t5")
model.eval()

note   = "James Smith, MRN482910, admitted with chest pain. DOB 03/22/1955."
prompt = (
    f"rewrite: [risk=0.72] [modality=asr] [version=1] "
    f"[trigger=cross_modal_linkage] note: {note}"
)

ids = tok(prompt, return_tensors="pt", max_length=256, truncation=True)
with torch.no_grad():
    out = model.generate(
        ids["input_ids"],
        attention_mask=ids["attention_mask"],
        max_new_tokens=200,
        num_beams=4,
        no_repeat_ngram_size=3,
        early_stopping=True,
    )
print(tok.decode(out[0], skip_special_tokens=True))

---

Input format

rewrite: [risk=FLOAT] [modality=MODALITY] [version=INT] [trigger=REASON] note: NOTE_TEXT

Token	Values
risk	0.0 to 1.0, cumulative re-identification risk from upstream scorer
modality	text, asr, image_proxy, waveform_proxy, audio_proxy
version	0 = no trigger fired, 1+ = number of pseudonym rotations
trigger	none, cross_modal_linkage, exposure_accumulation

---

PHI handling

PHI type	Replacement
Patient name	Consistent pseudonym, version-aware
DOB	[DOB-V{version}]
MRN	[MRN-V{version}]
Phone	[PHONE-V{version}]
Address	[ADDR-V{version}]

---

Training

Fine-tuned from t5-base (223M parameters) on 12,000 synthetic clinical note pairs with bracket tokens added as special tokens. Each pair was generated with injected fake PHI and a matching rewrite. Trigger scenarios (cross_modal_linkage, exposure_accumulation) were oversampled so the model learns version-bump behavior reliably. Trained on Apple M4 MPS.

Metric	Value
Base model	t5-base
Parameters	222,894,336
Training pairs	12,000
Epochs	10
Best val loss	0.1183
Format accuracy (raw)	100%
Format accuracy (normalized)	100%
Special tokens added	16 ([DOB-V0]–[ADDR-V3])
Trigger distribution	none: 49%, cross_modal_linkage: 38%, exposure_accumulation: 13%

Val loss curve: 0.2121 → 0.1521 → 0.1480 → 0.1438 → 0.1392 → 0.1330 → 0.1283 → 0.1226 → 0.1195 → 0.1183 (best every epoch).

Architecture: 12-layer encoder-decoder, 768 hidden dim, 12 attention heads, 3072 FFN dim.

---

Where it fits

DCPG Risk Scorer
      |
  Controller (policy decision)
      |
  DAGPlanner (remediation plan)
      |
  SynthRewrite-T5    <- this model
      |
  Rewritten safe clinical note

The risk score and trigger reason that condition this model come directly from the upstream controller and DCPG scorer. The version integer increments when FedCRDT-Distill or the controller flags a pseudonym rotation event.

---

Related

• phi-exposure-guard: full system
• dcpg-cross-modal-phi-risk-scorer: produces the risk score input
• exposureguard-policynet: policy enforcement layer upstream
• exposureguard-dcpg-encoder: graph encoder upstream
• exposureguard-fedcrdt-distill: federated risk and retokenization trigger
• exposureguard-dagplanner: remediation planner upstream
• streaming-phi-deidentification-benchmark: benchmark dataset
• multimodal-phi-masking-benchmark: PHI masking dataset with span annotations

---

Citation

@software{exposureguard_synthrewrite,
  title  = {ExposureGuard-SynthRewrite-T5: Stateful PHI-Aware Clinical Note Rewriter},
  author = {Ganti, Venkata Krishna Azith Teja},
  doi    = {10.5281/zenodo.18865882},
  url    = {https://huggingface.co/vkatg/exposureguard-synthrewrite-t5},
  note   = {US Provisional Patent filed 2025-07-05}
}

Trained on fully synthetic data. Not validated for clinical use.