Responsible Use of Mika AI Scribe

Information and guidelines for Canadian healthcare providers

Overview

Mika AI Scribe is reduces the time and effort you spend on charting, by drafting encounter notes and other documents based on encounter conversations.

Mika leverages two forms of cutting-edge artificial intelligence (AI) to deliver its capabililities:

  • Natural language processing (NLP)

  • Large language models (LLMs)

This article explains how these AI technologies work, including their data sources, limitations, and risks, and provides guidelines for responsible and safe use of Mika AI Scribe.


How Mika Works

It's important for you to understand how Mika -especially the AI powering it- works, so you can better understand the capabilities, limitations, and risks of the system.

Mika AI Scribe is a web-based software application that you can use on your mobile, tablet, laptop or desktop device.

When you start a Mika AI Scribe session, the application connects to the microphone on your device and begins securely streaming audio. The application uses an NLP service to transcribe the audio into a text transcript in real-time and then discards the audio.

This means that no audio recording of your conversation is made or kept by Mika AI Scribe.

When you stop a Mika AI Scribe session, the application sends the text transcript to an LLM service, along with instructions (i.e., a "prompt") and in some cases additional data (i.e., "context"). The LLM uses these inputs to generate a draft of your note according to the template you've chosen - such as SOAP notes, consult letters, or patient-friendly summaries.


Probabilistic versus Deterministic Systems

Traditional software is deterministic—given the same input, it always produces the same output.

AI-powered software like Mika, on the other hand, is probabilistic. It generates outputs based on likelihoods (i.e., probablities), which means:

The same input (e.g., transcript) can result in different outputs.

Try it for yourself: add the same note template to a Mika AI session several times. You'll see that the resulting notes are similar, but not exactly the same.

Outputs (e.g., notes) may appear reasonable and polished, but may still be incomplete or inaccurate.

You play a critical role in reviewing the notes Mika generates and double-checking them for both completeness and accuracy.

Mika does not "understand" clinical context the way a you do.

When Mika is drafting content for you to use in your notes, it is not using clinical context or reasoning the way you do.

Mika's AI is trained on large datasets and patterns within them to "predict" what words and content it should include in its draft, given the transcript and prompt for that session.

This technology is very similar to the technology that suggests the next word for you in your email or text messaging app (just at a MUCH bigger and more complex scale).

If you think of Mika as a "next word predicting machine" -albeit a very robust one- it helps you understand the importance of your review and oversight.


Deep Dive: Speech-to-text

In this step, the audio stream is transcribed in real-time, using natural language processing (NLP).

📥 Input

  • Real-time audio from your device's microphones, which can be complicated by accents, medical terminology, overlapping speech (i.e., when you and your patient talk at the same time), and varying levels of background noise.

📤 Output

  • A stream of timestamped text approximating what was said.

The Mika AI Scribe application does some post-processing on the output to create the finalized text transcript, which is then sent to the LLM along with the prompt and any additional context to draft your notes.

⚙️ How It Works (Simplified)

  1. Acoustic Modeling: Converts sound waves to phonemes (i.e., units of sound in a given language).

  2. Language Modeling: Uses probabilities of word sequences to "guess" what was most likely said.

  3. Vocabulary Mapping: Compares sounds against its database of words and phrases (We customize this further to include Canadian-specific medical terminology like drug names).

  4. Decoding & Post-processing: Outputs a transcription stream optimized for speed and readability.

🧠 What It’s Trained On

  • The underlying NLP model Mika uses is trained on a vast body of spoken language data, including conversations, speeches read aloud, and call center data.

    • This data is representative of a broad range of demographic factors

  • When we do additional customization (i.e., fine tuning), we use open source and/or synthetic data

⚠️ Common Limitations and Errors

Limitation
Example

Misheard words

“No known drug allergies” → “No known dry calories”

Homophones

“Their” vs “They’re” vs “There”

Overlapping speech

Can struggle to capture words when both doctor and patient speak

Medical terms

May struggle with less common clinical terms unless explicitly trained

Context loss

Doesn’t understand meaning — just sound and word patterns

🩺 Risks

  1. The transcript quality influences the note quality, so problems with the transcript can lead to problems, including errors, in the note.

  2. Patient quotes, including symptom information, can be distorted, influencing diagnosis and treatment decisions.

Tips and tricks for getting high quality transcripts

See our guides for troubleshooting microphone issues and optimizing your conversation to get great transcripts with Mika.

Microphone IssuesTranscripts

Deep Dive: Large Language Model

After transcription, Mika uses a large language model (LLM) to generate clinical note content based on the transcript.

📥 Input

  • A prompt that includes:

    • The text transcript from the previous step

    • Instructions about the patient or note structure (e.g., “generate a SOAP note”, “format each section into bullet points and write in a concise manner”)

    • Additional context (e.g., previous appointment notes) can be included if you're using our "Enhance your note" features. Learn more here.

📤 Output

  • Draft content for the various sections of the selected note template: e.g., content for the Subjective, Objective, Assessment, Plan sections of a SOAP note.

The Mika AI Scribe application does some post-processing on the output including filtering and validating the content, and then injects it into the Mika AI suggestion placeholders (i.e., the pink "jellybeans") in your note template.

⚙️ How It Works (Simplified)

The LLMs Mika uses belong to a family of AI called transformer-based neural networks. These LLMs "learn" language by analyzing patterns in massive datasets. Each generated word is selected based on:

  • The prompt (here, that refers to the transcript, instructions, and any additional context)

  • Probabilities derived from the model's training process

🧠 What It’s Trained On

  • The LLM Mika uses are trained on broad datasets including web pages (e.g., Wikipedia), books, papers, scientific literature, and dialogue and conversation data.

  • As with the NLP process, when we do additional customization (i.e., fine tuning), we use open source and/or synthetic data

⚠️ Common Limitations and Errors

Limitation
Example

Hallucinations

Fabricates plausible but incorrect clinical facts (e.g., mentions symptoms that weren't said)

Omission

Leaves out key findings or patient concerns

Bias

May reflect stereotyped assumptions about gender, race, age

Lack of reasoning

May generate notes that sound good but are logically flawed

Flattening

Polishes or flattens patient voice, removing subtle but important context

Inconsistency

If given slightly different prompts, outputs may vary in unexpected ways

🩺 Risks

  1. LLMs can fabricate information (hallucination), which if not caught, can introduce inaccuracies into your patient's' record.

  • LLMs may leave out relevant information, which if not caught, can introduce inaccuracies into your patient's record.

  • AI-generated content may appear reasonable and polished, but it must not be mistaken for authoritative documentation.


Guidelines for responsible, safe use of Mika AI Scribe

✅ You are the human-in-the-loop

Always review, correct, and sign off on any content drafted by Mika. Due to the probabilistic nature of the AI powering Mika, errors will occasionally occur (despite the guardrails we've put in place). You remain responsible for the information that ultimately ends up in the patient's record.

✅ Mika isn't a search engine, decision support tool, or a calculator

Mika helps document your thinking—it cannot do the thinking for you. Mika's output is often shockingly good; however, asking Mika to answer questions, come up with diagnoses and treatment plans, or perform calculations are NOT supported use cases. This means we haven't tested and evaluated Mika for accuracy, reliability, or safety for these kinds of scenarios.

✅ Flag and report issues

Your feedback helps us identify issues with Mika's accuracy, reliability, fairness and overall performance. We regularly review your ratings and comments. Please flag and report issues using the help button found on the session details page (more on that here) or simply Contact Us.

If you'd like to become more involved, we're always looking for providers to join our advisory group. Just contact us if you're interested.

✅ Optimize your transcripts

The quality of the transcript directly impacts the quality of the notes Mika generates.

See our guides for troubleshooting microphone issues and optimizing your conversation to get great transcripts with Mika.

Microphone IssuesTranscripts

✅ Remember that we're always here to help

Use us as a resource for learning more about the AI that powers Mika, as well as how to get the most out the application for your specific practice while still using Mika in a responsible way. We're easy to get a hold of and are happy to jump on a quick call to dig into your questions, issues, and ideas.

Last updated

Was this helpful?