Building Ambient AI Medical Scribes: Best Practices

5:11

Building Ambient AI Medical Scribes: Best Practices

AssemblyAI 15.12.2025 397 просмотров 12 лайков

Machine-readable: Markdown · JSON API · Site index

Смотреть на YouTube

Поделиться Telegram VK Бот

Транскрипт Скачать .md

Анализ с AI

Описание видео

🚀 Best practices for building ambient AI scribes: https://www.assemblyai.com/docs/medical-scribe-best-practices?utm_source=youtube&utm_medium=referral&utm_campaign=tutorials&utm_content=mart_ambient_scribe Learn how to build a simple AI ambient medical scribe in Python using AssemblyAI. In this tutorial, we start with a basic transcription request in under 30 lines of code and incrementally add features that matter in real clinical workflows. You’ll see how to improve medical transcription accuracy with key terms, label speakers by role (doctor vs. patient), generate SOAP notes using LLMs, and apply PII redaction to protect patient privacy. We’ll also cover how to delete transcription and LLM data after processing so your application meets strict data retention and privacy requirements. By the end, you’ll have a practical foundation for building a production-ready ambient medical scribe. Timestamps: 00:00 – Intro: Why accuracy and privacy matter for AI medical scribes 00:40 – Building a basic transcription pipeline in Python 01:01 – Speaker identification (doctor vs. patient) 02:18 – Generating SOAP notes with LLMs 02:56 – PII redaction and data privacy 03:40 – Deleting data and retention best practices 04:57 – Final recap ▬▬▬▬▬▬▬▬▬▬▬▬ CONNECT ▬▬▬▬▬▬▬▬▬▬▬▬ 🖥️ Website: https://www.assemblyai.com 🐦 Twitter: https://twitter.com/AssemblyAI 🦾 Discord: https://discord.gg/Cd8MyVJAXd ▶️ Subscribe: https://www.youtube.com/c/AssemblyAI?sub_confirmation=1 🔥 We're hiring! Check our open roles: https://www.assemblyai.com/careers ▬▬▬▬▬▬▬▬▬▬▬▬▬▬▬▬▬▬▬▬▬▬▬▬▬▬▬▬▬▬▬▬▬

Оглавление (7 сегментов)

Intro: Why accuracy and privacy matter for AI medical scribes

If you're building an AI ambient medical scribe, you've probably realized that your transcription layer can make or break your app. You need accuracy on industry terms, privacy that you can guarantee your customers, and you need your app to actually fit the real world scenarios and situations that your users are going to be in. That's why today I'll be building a very simple AI medical scribe and along the way I'll share some tips and tricks so that you can see how Assembly might be able to help you build the best AI medical scribe. So let's dig right in. What we have here is a very simple

Building a basic transcription pipeline in Python

transcription request in less than 30 lines of Python, but I'm going to add more features step by step and eventually this will be a more fleshed out AI medical scribe. This transcription request works as is. So why don't we just give it a quick try.

Speaker identification (doctor vs. patient)

So the transcription looks good. But now we want to add speaker labels so that we can differentiate between the doctor and the patient in the discussion. All right. So now we see the transcript separated by A and B. But if we use speech understanding, we can do speaker identification. Since we know that the roles in the conversation is doctor and patient, we can tell the API to label the transcript with their roles instead of just generic A and B. So here we specify speaker identification and we say speaker type role, doctor and patient. Let's give it a try. There we go. So now we see the doctor and the patient and we can differentiate between their role in the transcript itself. If I'm not wrong, the doctor prescribes Tremodol, but it is written in a way that isn't capitalized. So to fix that, I can add Tremodol to the key terms prompt, which basically biases the model to help transcribe it in the way that we specify. And now we see that the T in Tramodol is capitalized. Our transcription looks good. So why don't we generate some SOAP

Generating SOAP notes with LLMs

notes based off of this consultation as a summary for the doctor. To do this, I'll use LLM gateway which is an interface for various large language models which you can call with your assembly AI API key. So we'll add our prompt and then we'll make our request to LLM gateway. All we do here is pass the prompt in as a user message along with the transcript and then we call the request in a post request and we can print out the response. Let's try. All right. So here we can see the SOAP notes being generated.

PII redaction and data privacy

All of that looks good. But in a medical context, it's important that we delete the data after we make a request so as to preserve the patient's privacy. To redact this personal information in the transcript, we can add PII reduction, which basically ensures that any of that personal information doesn't appear in the transcript and therefore in the soap notes that we generate. So here we can see that the patients and the doctor's names are redacted according to our settings in PII. All right. So, we have an accurate transcription with PII redacted and we're also generating soap nodes. All of this looks great, but I think the next

Deleting data and retention best practices

step is to delete the data from Assembly AI servers after the request is completed. In a production environment, you'd save the results on your database instead of having it on ours. And that way, you can be certain that we don't store any of your data on our end. To do this, we just have to add a delete request at the end of our script. So each of these lines put in a delete request for the transcript and then for the LLM gateway response. And just to prove that we really do delete the data off of our servers, we'll make one final get request and print the result so that we can verify that it's deleted. All right, here you can see that the audio URL has been deleted by user. The text is empty and just states deleted by user. And so we can really tell that this transcript request was deleted from our servers. In case you were curious, because we used an audio URL from our own S3 bucket, that audio file isn't stored on our servers and is immediately deleted upon a delete request as well. We also offer a time to live on these requests such that they are automatically deleted after a set period of time. If this is something that you're interested in, please get in touch with our team. It looks like our AI ambient medical scribe

Final recap

is complete. I hope you enjoyed this video and if you have any questions, please feel free to reach out to our team. Bye.

Другие видео автора — AssemblyAI

Ctrl+V

Экстракт Знаний в Telegram

Экстракты и дистилляты из лучших YouTube-каналов — сразу после публикации.

Подписаться

Лучшие методички за неделю — каждый понедельник