Aama-AI

An intelligent maternal health companion for rural Nepal

Nepal’s maternal mortality ratio stands at 142 deaths per 100,000 live births — nearly twice the SDG target. The cause is rarely a lack of medicine. It is the absence of any intelligent system that can recognize danger, guide a woman to care, and prepare that care before she arrives.

Aama-AI is a five-module, offline-first AI platform trained on real Nepalese clinical data and deployed through the community health volunteers rural Nepal already trusts.

Explore the platform The problem

The Problem

Three delays, one silence

Medical researchers describe preventable maternal death through the Three-Delays Model (Thaddeus & Maine, 1994). In rural Nepal, all three delays operate simultaneously, unaddressed by any existing digital platform.

142

maternal deaths per 100,000 live births

4–6 hrs

typical distance to the nearest facility

15.6%

of pregnant women affected by perinatal mental health disorders

AI health platforms deployed in rural Nepal

Failure to recognize danger

Without a system to monitor vitals or symptoms, a woman has no way to know whether what she is feeling is ordinary discomfort or a life-threatening warning sign.

Failure to reach a capable facility in time

Even when danger is recognized, the nearest hospital is often hours away — and may not be the nearest one equipped to treat the specific emergency.

Failure to receive adequate care on arrival

Clinical teams learn of an emergency only when the patient is at the door, leaving no time to prepare staff, blood, or an operating theatre.

Five failures, addressed by nothing at once

Obstetric risk is not detected early
Clinical triage is unavailable
Emergency routing is absent
Mental health is invisible
Nutritional guidance is generic and inaccessible

Why Existing Solutions Fail

Every approach solves one fragment

Three decades of digital health interventions in Nepal have produced measurable but fragmented results. SMS interventions reach FCHVs but stay passive. Telemedicine helplines resolve cases only when a specialist happens to be available. Government systems report what services were delivered — not who is at risk. No platform has combined offline-first architecture with a maternal risk model trained on Nepalese clinical data. Aama-AI is the first.

Approach	Offline-first	Clinical intelligence	Locally-trained AI	Patient-facing
SMS / mHealth reminders Scheduled text messages and FCHV training	Yes	—	—	—
Dedicated pregnancy apps Stage-specific educational content	—	—	—	—
Telemedicine helplines Teleconsultation dependent on an available specialist	Partial	Partial	—	—
Government HMIS / DHIS2 Facility-level aggregate service reporting	—	—	—	—
Generic AI chatbots Unconstrained conversational health guidance	Partial	—	—	—
Aama-AI	Yes	Yes	Yes	Yes

The Platform

Five integrated AI modules

Each module reinforces the others, so a woman’s journey from symptom to risk assessment to emergency care is continuous — and available without a doctor physically present.

Module 01Addresses Delay 1

Maternal Risk Prediction

A community health volunteer logs a woman's vitals during a home visit. Within seconds, the model classifies her risk as Low, Moderate, or Severe.

An ensemble of Random Forest and XGBoost, trained on clinical data from a partner Nepalese teaching hospital, classifies risk from blood pressure, weight, temperature, oxygen saturation, haemoglobin, gestational age, and reported symptoms. SHAP explainability shows the volunteer exactly which vital triggered an alert — the model is never a black box. A Severe classification bypasses the chatbot entirely and fires an immediate alert to the doctor portal.

Vitals logged

BP, SpO2, haemoglobin, weight

Preprocessing

Imputation, scaling, class balancing

RF + XGBoost ensemble

SHAP-explainable

Low / Moderate / Severe

Severe → doctor alert

Module 02Addresses Delay 1

RAG-Constrained Triage Chatbot

A conversational assistant answers symptom questions in Nepali — grounded strictly in verified clinical guidelines, never in open-ended model knowledge.

The chatbot is constrained by Retrieval-Augmented Generation: guidelines from Nepal's Ministry of Health, WHO Nepal, and hospital protocols are chunked, embedded, and stored in a vector database. Every response is retrieved from this verified content, not generated freely, eliminating hallucinated medical advice. Structured intent classification detects danger-sign language and escalates to a human volunteer or doctor.

Woman asks a question

In Nepali, via mobile app

Retrieval from verified guidelines

MoHP, WHO, hospital protocols

Constrained response

Grounded, not generated freely

Danger sign detected → escalate

Module 03Addresses Delay 2

MCDM Hospital Routing

When an emergency is identified, the platform routes to the nearest hospital actually equipped to treat it — not simply the nearest one.

A TOPSIS multi-criteria decision algorithm scores reachable hospitals against weighted criteria — specialist availability, operating theatre, blood bank, ICU capacity, and distance — using a curated dataset of verified obstetric capability. Routing to a facility that lacks a blood bank or surgical capacity wastes the exact time a severe complication cannot afford.

Emergency identified

Severe risk or manual trigger

Hospital dataset queried

GPS + verified obstetric capability

TOPSIS scoring

Specialist, theatre, blood bank, ICU, distance

Route to nearest capable facility

Module 04Addresses Delay 1

Mental Health Screening

Perinatal depression and anxiety, affecting over 15% of pregnant women in Nepal, are surfaced through standardized screening — and always escalated, never counselled by the AI.

The chatbot conversationally administers the PHQ-4 and Edinburgh Postnatal Depression Scale (EPDS). A moderate-to-severe score triggers an immediate alert to the doctor portal and notifies the volunteer. The system deliberately stops short of offering counselling itself — its role is detection and escalation to a human professional, not treatment.

PHQ-4 / EPDS screening

Administered conversationally

Score evaluated

Threshold-based

Below threshold

Warm, supportive acknowledgment

At or above threshold → escalate

Module 05Addresses Delay 1

Nutrition & Anaemia Guidance

Dietary guidance grounded in foods that actually exist in rural Nepal, tied directly to each woman's haemoglobin trend.

A retrieval-augmented knowledge base delivers guidance using locally available foods — dal bhat, leafy greens, sesame seeds, dried fish, fortified flour — rather than generic dietary advice. Haemoglobin is logged as a vitals field, feeding both this module and the risk prediction model, so anaemia risk directly influences a woman's overall classification.

Haemoglobin logged or diet query

Retrieval from nutrition knowledge base

Locally available foods

Personalized guidance in Nepali

Feeds risk prediction model

Haemoglobin as a shared feature

How It Works

One journey, three interfaces

The five modules operate as a single continuous workflow across three purpose-built interfaces — two of them working entirely without internet, syncing only once a connection is available.

Android · Offline-first

Woman's App

Used directly by pregnant women who own a smartphone — logging vitals, querying the triage chatbot, and receiving risk alerts on their own device.

Android · Offline-first

FCHV App

Carried by Female Community Health Volunteers on their own device. Operated on a woman's behalf during home visits, reaching those without a personal smartphone.

Web · Online

Hospital Admin Portal

Used by doctors and hospital staff over the facility's internet connection — a real-time dashboard of incoming risk alerts, patient history, and emergency routing.

A woman's journey, step by step

1
Vitals are logged
Works offline
A woman with a smartphone logs her own symptoms and vitals. Where she has none, an FCHV visits and logs the same data on the FCHV app instead.
2
Risk is classified instantly
Works offline
The risk engine classifies Low, Moderate, or Severe on-device, with no connection required.
3
Guidance or emergency routing follows
Works offline
Non-severe cases get chatbot guidance grounded in verified clinical guidelines. A Severe classification calculates the nearest hospital with the required obstetric capability, using the last-synced hospital dataset.
4
The device reconnects
Syncs on reconnect
As soon as the woman's phone or the FCHV's device regains signal, vitals, risk history, and any alerts sync automatically to the cloud backend.
5
The hospital sees it first
Online
The hospital admin portal updates in real time. Doctors see the risk history and emergency alert before the patient arrives, with time to prepare a theatre, blood, or specialist staff.

Validation & Status

Where the platform stands today

Aama-AI is a functional prototype validated against defined performance targets, not a concept. This is an honest account of what is proven, what is in progress, and what remains a target.

Risk prediction model

Ensemble in training against a 76% accuracy / 0.71 recall logistic regression baseline; target ≥82% accuracy, ≥0.85 recall on severe cases

In Progress

RAG triage chatbot

97% correct handling across 200 clinical test prompts, zero hallucinated claims — exceeds the ≥90% intent accuracy, <2% hallucination target

Completed

MCDM hospital routing

Hospital capability dataset curated with clinical-expert-weighted criteria; validation against 50 emergency scenarios underway, targeting ≥95% routing accuracy

In Progress

Mobile app & doctor portal

Functional offline-first app and web monitoring dashboard operating end-to-end on cloud infrastructure

Completed

Ethical approval & field pilot

Institutional review submitted and under active review; a 20–30 woman field pilot is planned pending clearance

Target

Roadmap & Impact

From pilot to provincial scale

Based on comparable mHealth interventions across South Asia, Aama-AI targets a 10–15% reduction in preventable maternal complications within its deployment areas over the long term.

2026

Localized field pilot

20–30 pregnant women in a semi-rural community, deployed through existing FCHV networks, pending ethical clearance.

2027–28

Multi-district validation

Expanded deployment and continued model retraining as clinical data grows across Bagmati Province.

2029

Provincial-scale operation

Scaled deployment leveraging the existing health volunteer network to minimize infrastructure cost.

Aligned with global development goals

Good Health & Well-Being

Gender Equality

Reduced Inequalities