SAHAY: Multimodal, Privacy-Preserving AI for Suicide Risk Detection and Intervention in India

Salam Michael Singh, Manik Inder Singh Sethi, Suresh Bada Math, Tanmoy ChakrabortyIJCAI 2025 · DOI

Abstract

India faces a significant public health challenge with suicide being the leading cause of death among the 15–29 age group. Despite the scale of the crisis, a 75-80% treatment gap persists due to professional shortages and social stigma. This paper proposes SAHAY, a multimodal, privacy-preserving AI framework designed to bridge this gap. SAHAY integrates diverse data sources—social media, electronic health records, and wearable sensors—to provide culturally adaptive risk detection. Crucially, it connects these insights to India’s national tele-mental health service (TeleMANAS), enabling a tiered intervention strategy that "closes the loop" from detection to clinical care.

Introduction: The Suicide Crisis in India

In 2021, India reported over 164,000 suicide deaths, highlighting a critical need for scalable mental health interventions. The current landscape is marred by a severe shortage of professionals—only 0.75 psychiatrists per 100,000 people—and significant cultural barriers to seeking help. While AI-driven suicide prevention tools exist, they often suffer from fragmentation, lack of cultural context, and poor integration with existing healthcare infrastructures. Most models are optimized for English-speaking populations and fail to account for the unique linguistic and symptomatic expressions of distress prevalent in the Indian subcontinent.

The SAHAY Framework

SAHAY (meaning "help" or "support" in several Indian languages) is designed as a modular, "plug-and-play" architecture aligned with India’s Digital Public Infrastructure (DPI) principles. The framework moves beyond passive risk prediction to active intervention management.

The SAHAY end-to-end pipeline, from multimodal data sources to tiered intervention and geospatial outreach.

The system architecture consists of four primary pillars:

  1. Multimodal Data Integration: SAHAY aggregates data from three streams: social media (textual patterns), Electronic Health Records (EHR) for historical context, and Ecological Momentary Assessments (EMA) from wearables, tracking physiological markers like heart rate variability and sleep patterns.
  2. LLM-Based Risk Detection: Utilizing Large Language Models (LLMs) such as BLOOM, mT0, and India-specific models like OpenHathi and Airavata, the system is fine-tuned to recognize distress markers across 22+ Indian languages.
  3. Tiered Intervention Pipeline: Risk is categorized into Low, Medium, High, and Critical levels. This allows for efficient resource allocation: low-risk cases receive AI-assisted support, while higher-risk individuals are seamlessly transitioned to human counselors and psychiatrists via the TeleMANAS network.
  4. Geo-spatial Crisis Mapping: By identifying "suicide hotspots," the framework assists in deploying physical crisis outreach teams to underserved regions.

Data and Cultural Context

To ensure robust performance, SAHAY leverages several validated datasets, including MIND-LAMP for wearable sensing data, SHARP for longitudinal patient patterns, and established social media discourse datasets. Clinical instruments such as PHQ-9 (depression), GAD-7 (anxiety), and C-SSRS (suicide severity) scales are used for training and validation.

A core innovation of SAHAY is its focus on cultural nuance. Rather than simple translation, the models are trained to understand collectivist social structures and the "somatic" expression of distress—where emotional pain is often reported as physical symptoms (e.g., headaches or fatigue). This ensures that the AI can detect subtle red flags that Western-centric models might overlook. Furthermore, the framework employs a "Privacy by Design" approach, using federated architecture and differential privacy to protect sensitive user data while allowing for population-level analysis.

Evaluation and Future Outlook

As a proposal for the IJCAI-25 Special Track on AI and Social Good, SAHAY focuses on architectural feasibility and alignment with UN Sustainable Development Goals (SDGs). Success in upcoming deployments will be measured by risk detection precision across diverse languages, the efficiency of the "time-to-counselor" handoff, and longitudinal patient retention.

However, challenges remain. High-sensitivity models risk overwhelming the TeleMANAS network with false positives, and the system's ultimate success depends on the availability of physical medical resources in identified hotspots. Future work will focus on mitigating algorithmic bias and advocating for the integration of AI-assisted care into national health standards, potentially serving as a blueprint for other low-resource settings globally.

Conclusion

SAHAY represents a shift from fragmented AI tools to a comprehensive, integrated ecosystem for suicide prevention. By combining multimodal data with culturally sensitive LLMs and existing national health services, it offers a scalable solution to one of India’s most pressing public health crises.

Summary updated

BibTeX
@inproceedings{singh2025sahay,
  author = {Salam Michael Singh and Manik Inder Singh Sethi and Suresh Bada Math and Tanmoy Chakraborty},
  title = {{SAHAY: Multimodal, Privacy-Preserving AI for Suicide Risk Detection and Intervention in India}},
  booktitle = {International Joint Conference on Artificial Intelligence},
  year = {2025},
  doi = {10.24963/ijcai.2025/1096},
}
Figure