Inspiration

Doctors often spend more time documenting patient interactions than actually interacting with patients. Clinical notes are usually written after consultations, leading to delays, missing details, and cognitive overload for clinicians.

We were inspired to build HealthScribe to reduce this documentation burden by assisting doctors during consultations — not by replacing their judgment, but by helping them capture conversations accurately and consistently in real time.

What it does

HealthScribe is an AI-powered clinical documentation assistant that converts doctor–patient conversations into structured, editable SOAP notes.

It:

Captures real-time doctor–patient conversations using voice recognition

Transcribes natural speech into structured SOAP-format clinical notes

Highlights important clinical signals such as:

Worsening symptoms

Sleep disturbance

Functional impairment

Emotional distress

Delayed care indicators

Flags contradictory patient statements for clinician review

Supports multilingual input (Hindi/English with auto-translation)

Allows doctors to fully review, edit, approve, or reject all generated notes

Shares summaries with patients or pharmacies only after doctor approval

HealthScribe assists documentation only and does not provide medical advice, diagnosis, or treatment decisions.

How we built it

Frontend: React 18 with Vite, JavaScript (ES6+), CSS3, and Lucide React icons

Browser APIs:

Web Speech API for real-time speech-to-text transcription

MediaRecorder API for audio capture

AI / NLP (Rule-based):

Language detection and translation

Symptom and timeline extraction using heuristics

Clinical signal highlighting

Consistency and contradiction checking

Risk prioritization through rule-based logic

The system is designed so that all outputs remain editable and under the clinician’s control.

Challenges we ran into

Handling unstructured, conversational medical speech accurately

Avoiding over-interpretation of symptoms while still highlighting relevant clinical signals

Managing multilingual speech (Hindi–English code-mixing)

Designing a system that is helpful without crossing into medical decision-making

Ensuring transparency and trust in AI-generated clinical notes

Accomplishments that we're proud of

Built a working real-time voice-to-SOAP documentation pipeline

Successfully highlighted clinically relevant signals without automated diagnosis

Implemented contradiction detection to assist clinician review

Maintained strict human-in-the-loop control for all outputs

Designed a healthcare-safe AI system with clear scope boundaries

What we learned

Clinical AI must prioritize safety, explainability, and human oversight

Even simple rule-based NLP can provide high clinical value when designed carefully

Clear scope definition is essential when building healthcare technology

Multilingual healthcare tools significantly improve accessibility

Documentation support alone can have a meaningful impact on clinician workload

What's next for HealthScribe

Improved speaker separation (doctor vs patient)

More robust clinical signal detection with expanded rule sets

Integration with EHR systems (documentation-only workflows)

Support for additional languages

Enhanced clinician feedback loops for continuous improvement

Optional on-device processing for better privacy

Built With

Share this project:

Updates