MindfulMother: AI Powered Maternal Mental Health Support

Abstract

Maternal mortality remains a critical global health challenge, with India accounting for 12% of worldwide maternal deaths. Delayed recognition of danger signs and inadequate mental health support contribute significantly to preventable complications. This paper presents MindfulMother, a novel multi-agent artificial intelligence system designed to address these challenges through autonomous crisis detection and automated emergency response. The system implements a supervisor-based multi-agent architecture comprising four specialized domain agents Emergency, Hospital Finder, Mental Health, and Home Remedy orchestrated through a central coordinator with multi-signal intent classification. A key innovation is the five-priority cascading classification pipeline combining AI-driven semantic understanding with deterministic weighted keyword matching, ensuring zero-latency emergency detection independent of external API availability. Upon crisis detection, the system autonomously triggers telephony-based emergency response via Twilio Programmable Voice and SMS, transmitting the user\'s GPS coordinates to designated emergency contacts without requiring manual intervention. The architecture implements hierarchical memory management with dual-layer context session-based working memory with sliding-window summarization and persistent user profiles with episodic conversation summaries enabling personalized, context-aware interactions across sessions. Safety-critical pathways incorporate circuit breaker patterns for fault tolerance, comprehensive audit logging for clinical accountability, and a suspend-resume workflow pattern for human-in-the-loop confirmation. Integrated with a social support platform built on the MERN stack, the system provides community-driven peer support alongside AI-powered health assistance. Technical contributions include novel algorithms for multi-signal intent classification with safety override mechanisms, crisis severity assessment with dual-layer detection, sliding-window context summarization with LLM- based compression, and multi-signal content recommendation with temporal decay. The system demonstrates the feasibility of deploying safety-critical AI agents in healthcare contexts where system reliability directly impacts patient outcomes, offering a scalable architecture for life-saving maternal health intervention in resource-constrained settings.

Introduction

Maternal mortality remains a major global health issue, with India contributing significantly to the total deaths. Many complications arise due to delayed detection of danger signs, while postpartum depression often goes unnoticed. To address these challenges, MindfulMother is proposed as a full-stack web application that uses a multi-agent AI system for maternal health monitoring, crisis detection, and emergency response.

The system combines large language models with deterministic safety mechanisms to ensure reliable and fast crisis detection. It features a five-stage intent classification pipeline, bilingual (English–Hindi) support, and a dual-layer memory system for both short-term and long-term user context. In critical situations, it can automatically trigger emergency actions such as voice calls, SMS alerts, and GPS location sharing.

The architecture follows a three-tier design (frontend, backend, and data layer) with an additional AI services layer integrating tools like telephony and geolocation APIs. A supervisor-based multi-agent system coordinates specialized agents for emergency handling, mental health support, hospital search, and home remedies, ensuring controlled and safe decision-making.

Advanced algorithms are used for intent classification, crisis severity detection, emergency workflows, and context summarization. The system also includes human-in-the-loop confirmation, audit logging, and cooldown mechanisms for reliability and accountability.

Overall, MindfulMother demonstrates how AI-driven, safety-focused systems can improve maternal healthcare by enabling early detection, rapid response, and accessible support, especially in critical and resource-limited environments.

Conclusion

This paper presented MindfulMother, a multi-agent AI system designed to address critical gaps in maternal healthcare through autonomous crisis detection and automated emergency response. The system demonstrates that safety-critical AI agents can be deployed in healthcare settings when supported by structured architectural safeguards, including deterministic routing, layered validation mechanisms, fault-tolerant external service integration, and comprehensive audit logging. The proposed framework integrates a multi-signal intent classification pipeline combining structured LLM-based semantic analysis with deterministic bilingual keyword safety nets to ensure reliable crisis detection, including zero-latency fallback independent of API availability. An autonomous emergency workflow enables programmable telephony-based voice calls and GPS-enabled SMS alerts, reducing user burden during high-risk situations. The architecture further incorporates hierarchical memory for cross-session personalization, bilingual crisis detection with negator suppression to minimize false positives, and a multi-signal recommendation engine to strengthen community-based support within the same ecosystem. Although clinical validation, regulatory compliance, and expanded multilingual support remain areas for future work, MindfulMother illustrates how supervisor-based multi-agent coordination, circuit breaker fault tolerance, safety override mechanisms, and structured audit trails can enable responsible deployment of AI systems in safety-critical maternal healthcare contexts.

References

[1] World Health Organization. (2019). Trends in maternal mortality 2000 to 2017. WHO, Geneva. [2] World Bank. (2019). Maternal mortality ratio (modeled estimate, per 100,000 live births). World Development Indicators. [3] United Nations Population Fund (UNFPA). (2014). The State of the World’s Midwifery 2014. [4] World Health Organization. (2017). Depression and Other Common Mental Disorders: Global Health Estimates. [5] The Lancet Psychiatry. (2016). Maternal mental health and child health outcomes. [6] American Psychiatric Association. (2013). Diagnostic and Statistical Manual of Mental Disorders (DSM-5). Washington, DC. [7] Nature Medicine. (2019). Topol, E. High-performance medicine: the convergence of human and artificial intelligence. [8] JAMA. (2018). Rajkomar, A., Dean, J., & Kohane, I. Machine learning in medicine. [9] Multiagent Systems – Wooldridge, M. (2002). An Introduction to MultiAgent Systems. Wiley. [10] Artificial Intelligence: A Modern Approach – Russell, S., & Norvig, P. (2010). Pearson. [11] Autonomous Agents and Multi-Agent Systems. (2008). Jennings, N. R. Agent-based computing. [12] ACL. (2018). Papers on suicide ideation detection using NLP. [13] IEEE Intelligent Systems. (2017). AI-based decision systems in healthcare. [14] ACM Transactions on Information Systems. (2010). Adomavicius & Tuzhilin. Context-aware recommender systems. [15] ACM RecSys. (2015). Temporal decay models in social recommendation.

Copyright

Copyright © 2026 Pranita Lad, Ayush Kasare, Asmita Chavan, Kirtida Naik. This is an open access article distributed under the Creative Commons Attribution License, which permits unrestricted use, distribution, and reproduction in any medium, provided the original work is properly cited.