medirelay

“Closing the Medication Safety Loop.”

Comment

MediKiosk Relay: Don’t Assume. Verify.

Inspiration

MediKiosk Relay was inspired by a simple but important problem in crowded, multilingual OPDs: a patient may leave the consultation without fully understanding their medication instructions.

Doctors often have only a few seconds to explain a dose change, while elderly and low-literacy patients may struggle to recall their medical history or communicate comfortably in a language they are not familiar with. The dangerous part is that misunderstanding may not become visible until after the patient has already left the hospital.

We wanted to build something that doesn't simply give information to the patient, but checks whether the patient actually understood it.

That led us to the idea of a closed-loop OPD workflow.

What We Built

MediKiosk Relay is an AI-assisted, multilingual system designed to support the OPD workflow before consultation and verify medication understanding before discharge.

The workflow has six stages:

  1. Smart Pre-Consult Intake Patients provide their history through Hindi/English voice or touch-based interaction.

  2. AI-Built Medical Timeline Information from patient responses, prescriptions, and lab reports is structured into a timeline with source-linked information.

  3. Clinician Confirmation The doctor reviews and confirms high-risk information. The system does not provide a single "approve-all" action.

  4. Medication Safety Object The confirmed medication instructions are converted into a structured object containing fields such as:

  • Drug
  • Dose
  • Frequency
  • Duration
  • Timing/warning

  1. Voice Teach-Back Before discharge, the patient explains the medication instructions back to the system in Hindi.

  2. Deterministic Verification & Nurse Resolution The patient's response is compared field-by-field against the clinician-confirmed Medication Safety Object. If something differs, the system identifies the exact mismatched field and routes it for nurse intervention. The resolution is recorded in an audit trail.

The key principle is simple:

Don't assume the patient understood. Verify it.

How We Built It

The prototype combines several components into one workflow:

  • Hindi/English voice and touch-based intake
  • OCR for prescriptions and lab reports
  • Source-line-linked field extraction
  • LLM-based structured extraction with confidence scoring
  • Hindi automatic speech recognition for teach-back
  • A deterministic medication comparison engine
  • A rule-based clinical review signal
  • FHIR JSON export for interoperability
  • Audit logging for the complete workflow

We deliberately separated AI assistance from final verification.

The LLM helps structure information, while the critical medication comparison is deterministic. A clinician remains responsible for confirming high-risk information, and a nurse resolves mismatches rather than allowing the system to make an autonomous clinical decision.

What We Learned

One of our biggest learnings was that using more AI does not automatically make a healthcare system safer.

In a safety-sensitive workflow, we found it more important to clearly define what AI should do and what it should not do.

For example, an OCR or speech-recognition system can make mistakes. Instead of silently trusting those outputs, our workflow keeps confidence and source information visible and places human confirmation at critical points.

We also learned that a healthcare product needs to think beyond the model itself. Interoperability, auditability, human workflow, error handling, and deployment ownership are equally important.

FHIR compatibility was therefore included as a foundation for future integration rather than treating the prototype as an isolated application.

Challenges We Faced

1. Medical information extraction

Prescriptions and medical reports contain information that must be converted into structured fields without losing the original source context.

We addressed this using OCR with source-linked extraction and confidence scoring, followed by clinician confirmation for high-risk fields.

2. Hindi voice understanding

Free-form Hindi speech introduces ASR uncertainty. A wrongly transcribed number or frequency could potentially create a false mismatch.

For the prototype, we therefore use constrained prompts and show the raw transcript before verification so that errors remain visible and correctable.

3. Avoiding unsafe clinical conclusions

A combination such as elevated HbA1c/glucose and a refill gap should not automatically be interpreted as proof of non-adherence or deterioration.

Our system therefore treats this as a clinical review prompt, not a diagnosis or automated adherence verdict.

4. Defining the real deployment owner

Initially, "hospitals" was too broad as a target. We narrowed the operational focus toward the quality-and-patient-safety department or diabetes-clinic operations lead, where the audit trail and workflow metrics have a clearer purpose.

What's Next

The current prototype focuses deliberately on one complete, demonstrable workflow rather than claiming to solve every hospital use case.

Future work includes:

  • Support for multiple conditions
  • Broader medication safety workflows
  • Live EHR integration
  • Handwritten-prescription OCR
  • Larger-scale validation
  • Testing across more languages and patient populations

Our long-term goal is to extend the same closed-loop safety mechanism beyond diabetes OPDs to other situations where misunderstanding a high-risk medication instruction can have serious consequences.

The Core Idea

Healthcare communication should not end when the doctor gives an instruction.

It should end when the patient understands it — and that understanding has been verified.

MediKiosk Relay — Don’t Assume. Verify.

Accomplishments that we're proud of

What we learned

What's next for medirelay

Built With

Share this project:

Updates

Submission history