intelmedica/physician-sentences-1

---
license: cc-by-nc-4.0
language:
- en
task_categories:
- text-generation
- automatic-speech-recognition
tags:
- medical
- clinical
- physician
- synthetic
- asr-training
- sentence-generation
- clinical-documentation
pretty_name: "IntelMedica Physician Sentences v1"
size_categories:
- 100K<n<1M
dataset_info:
features:
- name: text
dtype: string
- name: category
dtype: string
- name: source_api
dtype: string
- name: term
dtype: string
- name: audience
dtype: string
splits:
- name: train
num_examples: 75534
- name: validation
num_examples: 16186
- name: test
num_examples: 16186
---

# IntelMedica Physician Sentences v1

Synthetic physician-specific clinical documentation sentences for training medical Automatic Speech Recognition (ASR) models. Part of the [IntelMedica](https://intelmedica.ai) open-source medical AI initiative.

## Overview

| Stat | Value |
|------|-------|
| **Total rows** | 107,906 |
| **Train** | 75,534 |
| **Validation** | 16,186 |
| **Test** | 16,186 |
| **Split ratio** | 70 / 15 / 15 (stratified by category) |
| **Language** | English |
| **Audience** | Physician |

## Category Distribution

| Category | Train | Val | Test | Total |
|----------|------:|----:|-----:|------:|
| soap_assessment | 13,928 | — | — | ~19,897 |
| hpi | 10,480 | — | — | ~14,971 |
| condition_drug | 10,075 | — | — | ~14,393 |
| ros | 6,971 | — | — | ~9,959 |
| discharge | 6,815 | — | — | ~9,736 |
| specialist | 5,591 | — | — | ~7,987 |
| surgical | 5,539 | — | — | ~7,913 |
| progress_note | 2,938 | — | — | ~4,197 |
| physical_exam | 2,923 | — | — | ~4,176 |
| consult | 2,921 | — | — | ~4,173 |

*19 categories total. Counts shown for train split; val/test follow same distribution.*

## Schema

| Column | Type | Description |
|--------|------|-------------|
| `text` | string | The generated clinical sentence |
| `category` | string | Clinical documentation category (e.g., sbar, hpi, soap_assessment) |
| `source_api` | string | Origin API of the medical term used in generation |
| `term` | string | The medical term the sentence was built around |
| `audience` | string | Target audience: `physician` |

## Data Sources

Medical terms were collected from 11+ authoritative APIs and databases:

| Source | Terms | Notes |
|--------|------:|-------|
| cross_source | 103,348 | Multi-API combined terms |
| rxnorm | 2,986 | NLM RxNorm drug names |
| dailymed | 792 | FDA DailyMed drug labels |
| snomed_ct | 768 | SNOMED CT clinical terms |
| loinc | 12 | LOINC lab test codes |

## Generation Pipeline

1. **Term collection** from 11 medical terminology APIs (RxNorm, SNOMED CT, NCI Thesaurus, MeSH, LOINC, DailyMed, HCPCS, FDA, CMS, plus curated nursing terms and 104K medical abbreviations)
2. **Quality cleaning** with 12 rules (deduplication, length filtering, encoding fixes, garbage removal) -- removed ~10% low-quality entries
3. **Template-based sentence generation** using Qwen 3.5 2B with audience-specific templates (physician clinical scenarios)
4. **Stratified splitting** into 70/15/15 train/validation/test by category

Full pipeline code: [intelmedica/med-speech-data-prep](https://github.com/intelmedica/med-speech-data-prep)

## Audio Versions

Audio versions (TTS-synthesized at 16kHz, multi-speaker) coming soon:

- `intelmedica/medical-tts-nursing-16khz`
- `intelmedica/medical-tts-physician-16khz`
- `intelmedica/medical-tts-general-16khz`

## Usage

```python
from datasets import load_dataset

ds = load_dataset("intelmedica/physician-sentences-1")
print(ds)
# DatasetDict({
# train: Dataset({features: ['text', 'category', 'source_api', 'term', 'audience'], num_rows: 75534})
# validation: Dataset({features: [...], num_rows: 16186})
# test: Dataset({features: [...], num_rows: 16186})
# })

print(ds["train"][0])
```

## Related Datasets

- [jfmdai/medical-speech-data-collections](https://huggingface.co/datasets/jfmdai/medical-speech-data-collections) -- Field directory of all medical speech datasets
- [jfmdai/nursing-sentences](https://huggingface.co/datasets/jfmdai/nursing-sentences) -- Original source (nursing)
- [jfmdai/physician-sentences](https://huggingface.co/datasets/jfmdai/physician-sentences) -- Original source (physician)
- [jfmdai/general-medical-sentences](https://huggingface.co/datasets/jfmdai/general-medical-sentences) -- Original source (general)

## Why `-1`?

This is **version 1**. Future versions will incorporate:

- Additional APIs (PubMed, RadLex, ClinicalTrials.gov)
- Accent diversity via voice cloning
- LLM-generated contextual clinical scenarios
- Real-world correction-based improvements from deployed ASR systems

## License

[CC BY-NC 4.0](https://creativecommons.org/licenses/by-nc/4.0/)

## Citation

```bibtex
@dataset{physician_sentences_1,
author = {Farooq, Junaid},
title = {IntelMedica Physician Sentences v1},
year = {2026},
publisher = {Hugging Face},
url = {https://huggingface.co/datasets/intelmedica/physician-sentences-1},
note = {Synthetic medical sentences for ASR training}
}
```

## Author

**Junaid Farooq, MD** / [IntelMedica LLC](https://intelmedica.ai) / Physician-Led Open-Source Medical AI

## Disclaimer

This dataset is for **research purposes only**. It is not a medical device, not Software as a Medical Device (SaMD), and not intended for clinical decision support. All data is **synthetic** -- no Protected Health Information (PHI) is present. Generated from publicly available medical terminology databases.