Cognizant Technology Solutions Corporation
Open
The Multimodal Triage Assistant
It will never diagnose you. It will make sure the person who can, has everything they need in ninety seconds.
Evaluation only · Open
Build a human-in-the-loop triage assistant for government hospitals, primary health centres, public health camps, company clinics, industrial-estate health units and campus health centres. It takes patient-described symptoms, uploaded reports and basic visual inputs, and turns them into a clean, structured triage note for a qualified professional to review. It organises. It flags urgency. It never plays doctor.
Overview
Outcomes - A structured triage note generated from messy, multilingual, multimodal patient input - Faster review for the doctor, nurse or medical officer, with missing information surfaced before the consult, not during it - Urgency signals highlighted and queues prioritised without any diagnosis being made - Referral notes for higher facilities prepared in advance instead of scribbled at the door - A demonstrable consent, anonymisation and audit trail that would survive an actual privacy review Overview The solution must work across India, where patient load, language diversity, specialist availability and digital maturity vary enormously between one facility and the next. The system is explicitly non-diagnostic. It organises information, highlights urgency signals and supports faster review. It must not prescribe treatment or replace a qualified professional. Suggested scenarios include outpatient queue triage, occupational-health screening in industrial estates, campus fever triage, maternal-health follow-up reminders, chronic disease check-in support, public health camp screening, and referral note preparation for higher facilities. The Story The queue starts forming at 6 AM. By 11 the doctor has seen sixty people and has roughly four minutes each. A patient arrives with a plastic bag containing three lab reports, two of them from 2023, one in a language the doctor does not read, and a description of the problem that begins with "it started around Diwali." Somewhere in that bag is the one detail that matters. Right now, finding it costs three of the four minutes. Give those three minutes back.
The brief
- The prototype should collect symptoms through text or voice, extract key details from sample medical reports, summarise timelines, identify missing information, and generate follow-up questions for a health worker, nurse, doctor or medical officer
- It may include OCR for lab reports, translation between English, Hindi and regional languages, risk-category tagging, queue prioritisation, referral preparation, and a reviewer dashboard
- Guidelines (read these twice): use synthetic or public sample data only. Do not use real patient records. Include a clear disclaimer that the solution is an educational prototype for triage-support purposes only
- Recommended technologies include OCR libraries, speech-to-text, translation, lightweight LLM summarisation, rules-based risk flags, and secure role-based access mockups
- Teams should demonstrate consent, minimal data retention, anonymisation, auditability and handoff to qualified medical staff. Any health-related output must remain advisory and reviewer-facing
Deliverables
- A working intake flow accepting symptoms by text and voice, in at least two languages, producing a structured triage note
- An extraction module that reads sample lab reports via OCR, pulls key values, and builds a timeline with gaps and missing-information prompts flagged
- A reviewer dashboard with risk-category tagging, queue prioritisation and a clear escalation-to-human path
- A referral note generator for handoff to a higher facility
- A privacy and responsible-AI pack: consent flow, anonymisation approach, retention policy, audit log, role-based access mockup, and the non-diagnostic disclaimer visible inside the product itself
Evaluation criteria
- Safety-first, explicitly non-diagnostic triage workflow — 20%
- Quality of information extraction and summarisation — 20%
- Multimodal capability — text, voice, OCR, image understanding — 15%
- India-wide facility relevance, accessibility, and human-review/escalation design — 30%
- Privacy, responsible AI controls and demo quality — 15%
