Kalusugan Triage: Smarter queues and safer referrals for barangay health centers

Problem Statement

Barangay health stations and rural health units are the front door of Philippine healthcare under the Universal Health Care Act (RA 11223). They are understaffed, and many rural municipalities have far fewer doctors than the WHO-recommended ratio. Patients line up from dawn on a first-come, first-served basis. A child with danger signs of dehydration or dengue may wait behind routine BP checks. Volunteer BHWs have limited training, records are often on paper, and referrals to hospitals are delayed or incomplete. Language barriers add to this, since many patients describe symptoms only in their local language. The result is late recognition of emergencies, overcrowded RHUs, and underused PhilHealth Konsulta benefits.

Proposed Solution

Kalusugan Triage is an offline-first Android/PWA app for BHWs and RHU nurses. A patient’s complaint is captured by voice or text in the local language. The AI transcribes and structures it into symptoms, duration, and vital signs entered by the BHW, then runs a protocol-based triage engine aligned with WHO/DOH guidelines (e.g., IMCI danger signs for children, maternal danger signs, dengue warning signs). The LLM never makes the final call. It only extracts information and explains. Each patient gets a color-coded priority (Emergency / Priority / Routine) and a queue number shown on a waiting room screen and sent by SMS, so patients can wait at home. Danger signs trigger an immediate alert and an auto-generated referral slip to the nearest hospital. Patients are also checked for Konsulta registration and nudged to enroll. Supervisors see daily load, wait times, and symptom trends such as dengue or ILI clusters.

Target Market

• Primary users: BHWs (several hundred thousand nationwide), midwives, nurses, and doctors in barangay health stations and RHUs.
• Buyers (B2G): Municipal and provincial health offices, DOH regional offices, Konsulta-accredited providers, and UHC integration sites. Development partners such as USAID, UNICEF, and WHO are potential funders.
• Size: Approximately 20,000+ barangay health stations and 2,500+ RHUs and health centers across 42,000+ barangays.

How AI Is Used

• Speech-to-text in Filipino languages (e.g., Whisper fine-tunes): Captures symptoms from patients who cannot write or who speak only Bisaya or Ilocano. Without it, intake stays slow and English-dependent.
• LLM information extraction: Turns free-form narratives (“tatlong araw nang nilalagnat, may pantal, masakit ang tiyan”) into structured symptom fields for the rule engine.
• Hybrid triage (rules plus ML ranking): DOH/WHO protocol rules guarantee safety-critical red flags. ML ranks non-urgent queues and predicts wait times.
• Forecasting and anomaly detection: Detects unusual spikes, such as fever with rash clusters after typhoons, for early outbreak signals to the municipal health office.
• Why it matters: Without AI, local-language intake and structured referrals are impossible at BHW skill levels.

PLDT / Smart / ePLDT / Rokid Angle

• Smart connectivity and SMS: Queue SMS, referral alerts, and offline sync over low-bandwidth Smart networks in rural areas.
• ePLDT cloud and data centers: Secure, locally hosted health data, which is important for sensitive personal information.
• PLDT Enterprise B2G: Bundles with connectivity for RHUs (aligned with DICT and DOH digital health) and telemedicine.
• Rokid AR (stretch): A BHW doing home visits sees danger sign checklists hands-free, or a remote doctor sees what the BHW sees during a teleconsult.

Demo Scenario (3 minutes for judges)

1. It is 7 AM at a Samar RHU and 40 people are in line. Mang Tonyo, a BHW, opens Kalusugan Triage.
2. A mother speaks in Waray/Bisaya about her 2-year-old: “Duha na ka adlaw nga nagsusuka, diri na nainom.” (“Vomiting for 2 days, can’t drink anymore.”)
3. The app transcribes and structures the complaint and flags “IMCI danger sign: unable to drink.” The card turns RED, the nurse’s tablet alerts, and the child jumps the queue.
4. A pre-filled referral slip is generated for the provincial hospital and sent by SMS.
5. Meanwhile an elderly patient for a BP refill is marked Routine and receives an SMS: “Your turn in about 90 minutes.”
6. The supervisor dashboard shows average wait time dropping and a mini-spike of fever cases in one barangay. Close with the line “Hindi ito doktor. Katuwang ito ng BHW.”

30-Hour Build Plan

Suggested stack: React PWA with IndexedDB (offline), Whisper/faster-whisper or a Filipino fine-tune, Gemini or SEA-LION for extraction, a Python rule engine, FastAPI, PostgreSQL, Grafana/Recharts, and an SMS gateway. Public data / APIs: WHO IMCI guidelines, DOH clinical practice guidelines and dengue advisories, the DOH Health Facility registry (NHFR), PhilHealth Konsulta provider lists, PSA OpenSTAT health statistics, Hugging Face Filipino speech datasets (e.g., Common Voice, FLEURS Filipino/Cebuano), and OpenStreetMap health facilities.

Risks & Privacy Notes

• Clinical safety: The tool does not diagnose. Red flags come from deterministic protocol rules, a health worker always confirms, and the UI states that it is decision support only. Clinical validation and DOH/FDA guidance are needed before deployment.
• ASR errors in dialects: The BHW confirms the extracted symptoms, with a fallback to a tap-based checklist.
• Data Privacy Act 2012: Health data is sensitive personal information, so the system requires explicit consent, encryption, local hosting, role-based access, and a Privacy Impact Assessment, and it must align with DOH-PhilHealth EMR standards.
• Clause 24 compliance: Only scripted, fictional patient cases are used during the hackathon. No real patient data goes into AI tools.

AI Use Disclosure (draft)

We use Whisper-based speech recognition for Filipino languages, an LLM (e.g., Gemini or SEA-LION) for symptom extraction, and simple ML for wait time prediction. Triage priority is decided by human-reviewed, protocol-based rules, not by the LLM. AI coding assistants helped scaffold the app, and test patient scripts were partly AI-generated. Team members reviewed all outputs and protocol mappings. No personal or health data was entered into AI tools.

Rubric Scoring