🩺 Med-Scribe & Alert
Real-time ambient listening tool for GPs — transcribes consultations, extracts symptoms, and alerts when a critical pattern is detected.
PPL Spark prototype · Witten/Herdecke University collaboration
Select a scenario and press Run to watch the alert panel respond as the conversation unfolds.
No EHR record loaded for this scenario.
Conversation
Alert Panel
Alerts will appear here...
Speak into your microphone — the system transcribes in real time and fires alerts when a critical symptom pattern is detected.
Supports English and German · Speaker diarization labels voices automatically
Separates Doctor / Patient voices (uses Deepgram)
Live Transcript
Alert Panel
Alerts will appear here...
🗺️ Product Journey — Med-Scribe & Alert
How the system fits into a GP consultation from door to alert.
The Problem Today
Patient describes symptoms
↓
Doctor listens & takes notes manually
↓
Diagnosis relies entirely on doctor's recall + experience
↓
Rare diseases are missed — avg. rare disease diagnosis takes 4–6 years
With Med-Scribe
| Step | What happens | Technology |
|---|---|---|
| 1 · Patient arrives | Doctor opens Med-Scribe, presses Start | macOS overlay / web UI |
| 2 · Conversation begins | Microphone captures ambient audio in real time | Deepgram Nova-3 STT |
| 3 · Transcription | Live rolling transcript, German & English, speaker-separated | Deepgram diarization |
| 4 · Symptom extraction | Every 30s: LLM reads new transcript, extracts symptoms, handles negations | GPT-4o / Claude |
| 5 · ICD-10 normalization | Symptoms mapped to standard codes ("Brustschmerz" → R07.9) | LLM NER |
| 6 · Red flag matching | Extracted symptoms compared against HPO/Orphanet/clinical guidelines | RAG vector DB |
| 7 · Alert | 🔴 CRITICAL alert appears on doctor's screen if ≥2 symptoms match a pattern | Alert UI |
| 8 · Doctor decides | Reviews alert, asks follow-up question, or escalates | Human in the loop |
| 9 · Session ends | Full transcript + summary saved | Markdown export |
Alert Escalation Logic
Symptom mentioned by patient
↓
Negation check ("no chest pain" → skip)
↓
Match against 10 Red Flag conditions
↓
1 match → 🟡 Watch
2+ matches → 🔴 CRITICAL
↓
Doctor sees alert in real time — without interrupting the conversation
What Stays Human
Med-Scribe is a smoke detector, not a diagnostician.
- The alert is a signal, not a diagnosis
- Doctor always decides what to do next
- No patient data is stored beyond the session (DSGVO by design)
- Works alongside existing PVS software — no integration required for MVP
📊 Competitive Landscape — Med-Scribe & Alert
May 2026 · Ambient AI scribes and clinical decision support tools for GP intake
Our Position
Key differentiators:
- 🇩🇪 German-language medical transcription (DSGVO-compliant, EU processing)
- 🔴 Real-time rare disease alerting (HPO/Orphanet matching) — no competitor does this
- 🏥 Designed for GP intake flow, not specialist/hospital workflows
- ⚡ Ambient — no manual input required
Ambient AI Scribes
The dominant and fastest-growing category. All focus on documentation efficiency — none perform real-time rare disease alerting.
| Product | Strengths | Gap vs. Us |
|---|---|---|
| Nuance DAX (Microsoft) | Market leader, Epic/Cerner integration | English-first, no alerting, US-centric |
| Abridge | Strong clinical accuracy, UPMC partnership | No diagnostic alerting, English-first |
| Nabla | European, GDPR-aware, multi-language | No rare disease alerting, limited German |
Diagnostic Decision Support (DDx)
| Product | Strengths | Gap vs. Us |
|---|---|---|
| Isabel DDx | Deep DDx engine, ICD-10 | Manual input required, not ambient |
| Ada Health | Consumer symptom checker, multilingual | B2C, not clinical workflow |
We are the only solution in the ambient + diagnostic alerting quadrant for German-speaking GPs.
Scenarios: MI (EN) · Sepsis (DE) · DVT→PE (EN) · Febrile Neutropenia with EHR (EN)
Live Session: real mic input · EN/DE · speaker diarization · same alert logic
LLM: BioMistral (HF) · GPT-4o (OpenAI) · Claude Haiku (Anthropic) · STT: Deepgram Nova-3