Medical transcription and local LLM: data compliance patient

#HDS framework and medical confidentiality

Medical confidentiality (Article L.1110-4 of the French Public Health Code) applies to any document containing identifiable health information, including audio transcripts. As soon as patient data is processed outside the premises of the practice or hospital, the hosting provider must be HDS-certified (ASIP/ANS framework). In practice, sending a consultation to a non-HDS cloud API—OpenAI Whisper, Claude, or any mainstream public platform—puts you outside the legal framework, even for a test.

The answer: never let the audio or transcription leave the scope you control. That is precisely what a local LLM medical transcription pipeline enables: the inference machine is in the office, the disk is encrypted, and the network is isolated. No hosting provider, no authorization to request.

#100% local architecture

The pipeline consists of four components, all local: audio capture → Whisper-large-v3 → LLM with SOAP template → export to the practice-management software (LGC) or the DMP.

Capture

Lavalier mic or cardioid USB mic placed between practitioner and patient. No Bluetooth (latency and compression). Record in mono WAV at 16 kHz.

Transcription

Whisper-large-v3 (OpenAI, open weights, MIT license) run via faster-whisper on a local GPU. Model loaded into memory; audio destroyed after transcription.

Structuring

Llama 4 Scout (17B-A2B, MoE) or Qwen3-14B served by Ollama. The LLM consumes the transcription and produces a SOAP (Subjective, Objective, Assessment, Plan) note.

Export

The report is inserted into the EHR (Weda, Medistory, AxiSanté, Doctolib Pro, Maiia) via clipboard, an HL7 CDA shortcut, or an API if available.

#Hardware and software requirements

GPU

RTX 4070 12 GB minimum, RTX 4080 16 GB or RTX 4090 24 GB recommended. Whisper-large-v3 uses ~3 GB of VRAM, while the 14B Q4_K_M LLM uses around 9 GB.

Mac alternative

A Mac mini M4 Pro with 48 GB of unified memory runs the full stack. Ideal for a professional practice: quiet, fanless when idle, and low-power.

Storage

LUKS-encrypted NVMe SSD (Linux), BitLocker (Windows Pro), or FileVault (macOS). 100 GB is enough: models ~25 GB, with the rest for temporary transcriptions.

Network

The inference workstation is on a VLAN separate from the patient Wi-Fi. No outbound Internet access in production—only during supervised updates.

Software

Ollama (≥ 0.5) to serve the LLM on http://localhost:11434, faster-whisper (Python) for transcription, and ffmpeg for audio conversion.

#1. Whisper-large-v3 on French consultations

faster-whisper is a CTranslate2 fork that is 4 to 5× faster than the reference Python implementation, at identical quality. It accepts the same models and exposes a simple API.

python -m venv ~/venv-medtranscript
source ~/venv-medtranscript/bin/activate
pip install faster-whisper==1.1.0 ffmpeg-python

# Téléchargement du modèle (téléchargé une fois, ~3 Go)
python -c "from faster_whisper import WhisperModel; WhisperModel('large-v3', device='cuda', compute_type='float16')"

The transcription code itself fits in a few lines. Note the language='fr' setting, which disables automatic detection and prevents drift at the beginning of a consultation.

from faster_whisper import WhisperModel
import sys, pathlib

AUDIO = pathlib.Path(sys.argv[1])
model = WhisperModel('large-v3', device='cuda', compute_type='float16')

segments, info = model.transcribe(
    str(AUDIO),
    language='fr',
    vad_filter=True,             # coupe les silences
    vad_parameters={'min_silence_duration_ms': 500},
    beam_size=5,
    initial_prompt=(
        'Consultation médicale en français. Vocabulaire clinique, '
        'noms de molécules, posologies. Termes courants : doliprane, '
        'amoxicilline, lévothyrox, ECG, IRM, NFS, CRP, HbA1c.'
    ),
)

transcript = '\n'.join(seg.text.strip() for seg in segments)
OUT = AUDIO.with_suffix('.txt')
OUT.write_text(transcript, encoding='utf-8')
print(f'Transcription : {OUT}')

On a RTX 4080, a 15-minute consultation is transcribed in about 90 seconds. The source audio file is deleted immediately afterward—the transcription is sufficient, and the audio creates unnecessary risk.

#2. LLM and SOAP template

The SOAP note is the de facto standard: Subjective (reason for visit, patient complaints), Objective (clinical examination, vital signs, additional tests), Assessment (diagnosis or hypotheses), Plan (prescriptions, upcoming tests, follow-up). The LLM is instructed to fill out this template strictly, without making anything up.

# Option 1 — Llama 4 Scout (multimodal, 17B MoE, ~10 Go en Q4)
ollama pull llama4:scout

# Option 2 — Qwen3 14B (excellent en français)
ollama pull qwen3:14b

# Option 3 — Mistral Magistral 24B si VRAM ≥ 16 Go
ollama pull magistral:24b

The system prompt enforces a strict framework: no invention, literal citation when uncertain, and French medical language.

Tu es un assistant médical chargé de structurer une transcription
de consultation au format SOAP. Tu écris en français médical clair.

RÈGLES IMPÉRATIVES :
1. N'INVENTE JAMAIS de symptôme, diagnostic, traitement ou posologie
   qui ne figure pas explicitement dans la transcription.
2. En cas d'information ambiguë, écris littéralement :
   "À préciser par le praticien : [extrait textuel de la transcription]".
3. Conserve les unités exactes (mg, mL, fois/jour). Ne convertis rien.
4. Pour toute posologie, cite mot pour mot la phrase d'origine
   entre guillemets dans la section Plan.
5. Ne formule pas de diagnostic définitif si le praticien ne l'a pas
   posé. Utilise "hypothèse diagnostique" ou "à confirmer par...".
6. Garde uniquement ce qui relève de la consultation. Ignore les
   éléments hors propos (conversations annexes, interruptions).

STRUCTURE DE SORTIE (Markdown, exactement ces sections) :

## Subjective
- Motif de consultation :
- Histoire de la maladie actuelle :
- Antécédents pertinents évoqués :
- Traitements en cours évoqués :

## Objective
- Examen clinique :
- Constantes mentionnées :
- Examens complémentaires cités :

## Assessment
- Hypothèse(s) diagnostique(s) :
- Diagnostics différentiels évoqués :

## Plan
- Prescriptions (citer textuellement) :
- Examens complémentaires demandés :
- Conseils donnés au patient :
- Suivi prévu :

À la fin, ajoute une section :

## Points à vérifier par le praticien
Liste des éléments douteux, manquants ou nécessitant validation.

#3. End-to-end pipeline

We chain transcription and then structuring through Ollama's local REST API. The script remains compact and auditable — about 40 lines, intentionally: less code means a smaller attack surface.

import sys, json, pathlib, requests, hashlib, datetime
from faster_whisper import WhisperModel

AUDIO = pathlib.Path(sys.argv[1])
MODEL_LLM = 'qwen3:14b'
SYS_PROMPT = pathlib.Path('system_prompt_soap.txt').read_text('utf-8')

# --- 1. Transcription locale
whisper = WhisperModel('large-v3', device='cuda', compute_type='float16')
segments, _ = whisper.transcribe(str(AUDIO), language='fr', vad_filter=True)
transcript = '\n'.join(s.text.strip() for s in segments)

# --- 2. Structuration SOAP via Ollama (localhost uniquement)
resp = requests.post(
    'http://localhost:11434/api/chat',
    json={
        'model': MODEL_LLM,
        'messages': [
            {'role': 'system', 'content': SYS_PROMPT},
            {'role': 'user', 'content': transcript},
        ],
        'options': {'temperature': 0.1, 'num_ctx': 8192},
        'stream': False,
    },
    timeout=600,
)
soap = resp.json()['message']['content']

# --- 3. Sortie + hash d'intégrité pour le journal d'audit
ts = datetime.datetime.now().isoformat(timespec='seconds')
sha = hashlib.sha256(soap.encode('utf-8')).hexdigest()[:16]
out = AUDIO.with_suffix('.soap.md')
out.write_text(
    f'<!-- généré le {ts} — modèle {MODEL_LLM} — sha256:{sha} -->\n\n{soap}',
    encoding='utf-8',
)

# --- 4. Effacement immédiat de l'audio et de la transcription brute
AUDIO.unlink()
print(f'Compte-rendu : {out}')

A 15-minute consultation completes the pipeline in under 3 minutes on RTX 4080. The .soap.md report is ready to paste into the LGC.

#4. Practice-management software integration

Three integration levels depending on your LGC:

1. 01
   Level 1 — Clipboard

   The script automatically copies the report (pyperclip). The practitioner pastes it into the LGC in two clicks. Compatible with 100% of LGCs, with zero integration on the vendor's side. This is the recommended starting point.
2. 02
   Level 2 — HL7 CDA shortcut

   The Markdown report is converted to CDA R2 (Clinical Document Architecture) by a pandoc script plus an XML template, then imported through the LGC's import documentaire function. Compatible with Weda, Medistory, and AxiSanté, which accept CDA.
3. 03
   Level 3 — Native LGC API

   Some vendors (Doctolib Pro, Maiia) expose an observation-insertion API. The pipeline pushes the report directly into the patient record identified by INS. This is the most convenient option but requires a vendor agreement and CPS authentication.

#Compliance and audit logging

Use of a medical writing assistance tool must be documented. At least three items in the patient record:

AI-assisted writing disclosure

A line at the bottom of the report: "Structured report assisted by a local AI model (Qwen3-14B, 2026-04-12 version). Validated by Dr. [...] on [date]." This statement provides traceability.

Access log

Each pipeline run is logged locally (who, when, which audio, report hash). The log is append-only, signed, and retained for 10 years (the patient-record retention period).

GDPR DPIA

Conduct an impact assessment (Article 35 GDPR) for the “AI-assisted transcription” processing activity. Being 100% local simplifies the assessment: no transfer outside the EU, no processor, single purpose.

Patient information

The waiting-room poster and the notice in the intake form must indicate the use of a transcription-assistance tool. The patient may refuse—plan an alternative workflow.

Encryption at rest

The entire disk is encrypted. Reports are stored in the patient's LGC folder, not on the inference workstation's file system. Purge temporary .soap.md files at the end of the day.

Supervised updates

Every model update (Whisper, LLM) triggers revalidation on a set of 10 anonymized test queries. Known regression bug = no deployment.

#Go further

This pipeline covers the everyday needs of a private practitioner or a small hospital department. Three natural extensions: industrialize multi-workstation deployment with Docker and a CPS-authentication reverse proxy, add a local RAG over HAS recommendations to assist with Assessment, and harden network security all the way to a complete air gap.