OpenMed onboarding for users in India¶

हिन्दी संस्करण

This guide shows an India-oriented, local-first de-identification setup: start with a policy that OpenMed ships today, de-identify synthetic Aadhaar- and ABHA-format identifiers, process a synthetic code-mixed Hinglish note, and select a compact Hindi checkpoint for a CPU-only machine.

For a complete synthetic Hindi and code-mixed Hinglish walkthrough, run the end-to-end example or follow the Chinese and Hindi notebook tour.

Install the local runtime¶

Create a virtual environment, then install OpenMed with the Hugging Face model dependencies:

python -m venv .venv
. .venv/bin/activate
python -m pip install "openmed[hf]"

Model download needs a network connection once. After the checkpoint is in the local Hugging Face cache, OpenMed inference runs on the machine and does not send the clinical note to a hosted API.

DPDP-oriented policy quickstart¶

The Digital Personal Data Protection (DPDP) framework is broader than text masking. Purpose and lawful processing, notices and consent where applicable, retention, access controls, incident response, contracts, and human governance remain organizational responsibilities. Consult the official DPDP Act, 2023 and DPDP Rules, 2025, and obtain appropriate legal advice. A software policy profile is one technical control, not a claim of DPDP compliance.

OpenMed ships the assist-only india_dpdp_act profile. It replaces direct identifiers, masks quasi-identifiers, keeps clinical concepts, requires the safety sweep, and disables both reversible IDs and retained mappings:

from openmed.core.policy import list_policies, load_policy

policy_name = "india_dpdp_act"
assert policy_name in list_policies()

policy = load_policy(policy_name)
assert policy.default_action == "replace"
assert policy.action_for("PERSON") == "replace"
assert policy.action_for("ID_NUM") == "replace"
assert policy.safety_sweep_mandatory
assert policy.keep_mapping is False
assert policy.reversible_id is False

print(policy.name, policy.default_action)

This code loads the shipped openmed/core/policies/india_dpdp_act.json profile. Its metadata records the official sources and makes clear that the profile is assist-only, not legal advice or an autonomous compliance determination.

De-identify synthetic Aadhaar and ABHA identifiers¶

UIDAI describes Aadhaar as a 12-digit number, while the Ayushman Bharat Digital Mission describes ABHA as a 14-digit number. The example uses these deliberately fabricated, display-form values:

• Synthetic Aadhaar-format value: 2467 7832 5484 (also passes OpenMed's Verhoeff checksum validator)
• Synthetic ABHA-format value: 91-0000-0000-0000

The india_dpdp_act profile automatically enables OpenMed's ABDM recognizer, which validates these Aadhaar and ABHA display forms and normalizes them to the canonical ID_NUM label. The small custom recognizer below only makes the fabricated person's name deterministic for this offline tutorial. The policy then controls the replacement action:

from openmed import deidentify

synthetic_note = (
    "Synthetic Hinglish note: Patient Asha Verma ka follow-up aaj hai. "
    "Aadhaar 2467 7832 5484 aur ABHA 91-0000-0000-0000 record mein hain. "
    "Patient ko do din se halka bukhar hai."
)

india_recognizer = {
    "case_sensitive": False,
    "deny": {
        "terms": [{"term": "Asha Verma", "label": "PERSON"}],
    },
}

result = deidentify(
    synthetic_note,
    method="replace",
    model_name="OpenMed/OpenMed-PII-Hindi-SuperClinical-Small-44M-v1",
    lang="hi",
    locale="en_IN",
    policy="india_dpdp_act",
    use_safety_sweep=True,
    custom_recognizer=india_recognizer,
    consistent=True,
    seed=707,
)

assert "2467 7832 5484" not in result.deidentified_text
assert "91-0000-0000-0000" not in result.deidentified_text
assert "Asha Verma" not in result.deidentified_text
print(result.deidentified_text)

The Hindi language hint selects Hindi-aware normalization and patterns. The compact Hindi model handles the code-mixed note, while the explicit local rules make the tutorial deterministic for the synthetic name and identifier formats. For a real deployment, derive custom rules from your documented data contracts, test direct-identifier recall on representative synthetic fixtures, and review every residual leak before release.

Run the complete example from a source checkout:

python examples/onboarding_india_dpdp.py

Low-RAM, CPU-only setup¶

Start with one of the small Hindi PII checkpoints already listed in OpenMed's model manifest. Raw FP32 weight size is approximately four bytes per parameter; the Python runtime, tokenizer, temporary tensors, and input length add overhead.

These small checkpoints map to OpenMed's Tiny device tier, whose release target is at most 350 MB resident RAM. Current tokenizer-script coverage marks only the 44M checkpoint as supporting Devanagari; the 33M and 66M checkpoints must not be selected for Devanagari notes without a new passing script audit. Treat the memory target as a release budget, not a guarantee for every Python environment; measure peak RSS on the exact machine and input lengths you will deploy. See Device Tiers and SLOs and Tokenizer Script Coverage.

For a constrained clinic workstation or startup VM:

1. Start with the 44M checkpoint, CPU execution, and batch_size=1 when notes can contain Devanagari. Consider the 33M checkpoint only for a tested, Latin-script-only Hinglish corpus.
2. Split long notes at sentence boundaries before batching. Keep adjacent sentences in order and never split a structured identifier across chunks.
3. Increase to batches of 2, then 4, only after observing enough peak-memory headroom with representative notes. Larger batches improve throughput but raise temporary tensor memory.
4. Reuse one loaded model or one BatchProcessor instead of constructing a loader for every sentence. Avoid running several model-worker processes on a low-RAM host.
5. Keep india_dpdp_act, the safety sweep, and the local identifier rules enabled when reducing model size. Re-run leakage tests before accepting a smaller checkpoint.

For multiple short notes, BatchProcessor(operation="deidentify") exposes the document batch_size control and reuses the underlying loader. See Batch Processing for the API.

Production checklist¶

• Keep inference and cached artifacts inside the approved device or network boundary; OpenMed does not add telemetry to the de-identification path.
• Do not retain result.original_text, raw entity values, or reversible mappings in logs or audit exports.
• Add locally relevant identifier patterns and synthetic regression fixtures; this walkthrough is not an exhaustive India identifier catalogue.
• Validate Hindi, Latin-script Hinglish, and code-mixed notes separately.
• Treat DPDP governance and ABDM participation requirements as independent of the selected OpenMed policy profile.