Implementing AI in Public Healthcare: Data, Ethics & System Readiness | India AI Impact Summit 2026

Contents

Executive Summary

This panel discussion examines the real-world implementation of AI tools in India's public healthcare systems, specifically focusing on maternal health, TB screening, and community health worker support. The core insight is that algorithmic sophistication means nothing without addressing data representation gaps, infrastructure constraints, and the broader healthcare ecosystem—emphasizing that effective AI deployment requires grounding in local context, existing workflows, and measurable health outcomes rather than technological capability alone.

Key Takeaways

Technology is Not the Bottleneck—Context Is: Successful AI in healthcare requires deep knowledge of local infrastructure, supply chains, protocols, staffing, and cultural norms before any model is deployed. Without this groundwork, even sophisticated AI fails to translate to health outcomes.
Representativeness Requires Active Intent: Underrepresented populations will remain underrepresented unless intentional, non-extractive efforts are made to include them in data collection. Data solidarity and legislative protections against misuse are as important as data gathering itself.
Human Oversight Is Non-Optional: LLM-based tools in healthcare must be bounded (RAG-based, protocol-grounded) and monitored by humans. "Tech plus touch" is not a nice-to-have—it's a requirement for safety.
Continuous Local Validation Beats Centralized Rollout: Rather than scaling a single solution nationally, design platforms that allow local teams to test, evaluate, and adapt solutions to their specific populations and constraints (PMAN framework / massively online AB testing).
Voice & Language Accessibility Are Equity Issues: Solutions that assume smartphone ownership, app literacy, and English/Hindi proficiency will exclude the poorest and most vulnerable. Voice-based interfaces and multilingual support are technical requirements, not luxuries.

Key Topics Covered

Data Representation & Bias: Mismatches between training data (US/China) and deployment contexts (India); differential performance across populations (23% higher false negative rates for pneumonia in darker skin tones; melanoma detection errors)
Infrastructure Gaps: Computational resource limitations, human capacity building, offline-capable tools, multilingual accessibility, and digital literacy barriers
Translation Gap: The disconnect between predictive models and actionable health outcomes—diagnostic capability without therapeutic power
Community Health Worker (CHW) Integration: LLM-based chatbots (WhatsApp, voice-based) designed to support ASHAs, ANMs, and other frontline workers
Protocol-Driven AI: Using retrieval-augmented generation (RAG) to ground chatbots in validated government protocols rather than free-form LLM outputs
Heterogeneity in India: Geographic, socioeconomic, cultural, and linguistic diversity requiring locally tailored solutions rather than one-size-fits-all deployments
Data Ecosystems & Privacy: Data solidarity, DPDP Act implications, health data exchanges, and shifting incentives around data collection
Evidence Generation: Moving beyond RCTs toward "massively online AB testing" (PMAN framework) for continuous validation in real-world settings
Last-Mile Populations: Focus on underrepresented groups—migrant mothers, tribal populations, religious minorities, adolescent mothers—whose data is often absent from datasets

Key Points & Insights

Data Bias is Foundational: Over half of clinical AI models use US or China training data, creating systematic failures when deployed in India. Example: 23% higher false negative rates for pneumonia detection in certain populations due to training data mismatch.
Grounding Prevents Hallucination: Successful implementations (Arman's ANM chatbot, Kushi Baby's ASHA chatbot) use RAG architecture, limiting responses to validated protocols and government guidelines rather than free-form LLM outputs. This ensures safety and accuracy.
Context Must Precede Technology: Before deploying a chatbot or AI tool, ground-truthing is essential—understanding medication supply chains, follow-up infrastructure, clinic availability, trained staff, and state-specific protocols. Generic tools fail without this foundation.
Voice-Based Interaction Solves the Digital Divide: Non-smartphone voice interfaces (accessible via phone calls, with auto-language recognition) bypass the assumption of app-based digital literacy and smartphone ownership—critical for reaching true last-mile users.
Human-in-the-Loop is Non-Negotiable: Even high-accuracy chatbots (97–98% satisfaction) require human oversight for edge cases, complex queries, and safety-critical decisions. "Tech plus touch" is essential.
Heterogeneity Demands Local Validation, Not Global Models: India's internal diversity (geographical, linguistic, cultural, socioeconomic) is so pronounced that a single national AI solution will fail in some regions. Platforms for local evaluation and deployment are needed.
Data Gaps Are Structural and Gendered: Minorities and underrepresented groups (migrant mothers, tribal populations, Muslim women, adolescent mothers) are absent from datasets entirely. Women's health data itself is historically male-biased. Active, intentional work to digitize and include these populations is prerequisite.
Procurement & Benchmarking Are Unsolved: No standard exists for comparing AI health solutions across implementations. Without benchmarking mechanisms, procurement becomes paralyzed and decision-makers cannot identify which solution works for their context.
Incentives Drive Data Quality: Reframing data reporting as opportunity rather than punishment (e.g., high TB mortality rates trigger investigation, not blame) encourages honest reporting and identifies where communities are being missed.
Language and Colloquial Terminology Matter Enormously: Medical terms translated directly (e.g., "diabetes") may not be understood by end users who use colloquial terms ("sugar"). Anemia ("kami" in Hindi) requires culturally grounded terminology to be effective.

Notable Quotes or Statements

"If there's no medication supply or no follow-up infrastructure or no accessible clinic or no trained doctors, we still can't act on it. And what that results in is a diagnostic gap without any therapeutic power." — Opening remarks on the translation gap

"The chatbot doesn't draw from the worldwide web. It draws answers only from the universe of the protocol that we created, which has been vetted so thoroughly at every level." — Dr. Aperna Hay (Arman), on protocol-grounded AI design

"Tech plus touch." — Dr. Aperna Hay, summarizing the essential combination of AI and human oversight

"If we wait for models to be built until we have all the data, we will never build the models. What we need instead is a way to test these models continually, look at where they are going wrong, and use that to retune and retrain them." — Dr. Anurag Aggarwal (Ashoka University), on iterative local validation vs. waiting for perfect datasets

"India itself is so heterogeneous that any one central program for the around the country is not going to work well. What you need are platforms to locally evaluate your problems." — Dr. Anurag Aggarwal, on the heterogeneity problem

"Data is not something to be afraid of but an opportunity that we can build on." — Dr. Mitun Kumar Mitra (IIT Mumbai), reframing data reporting mindset

"Even the way that the ASHA is able to ask questions from the chatbot—they will not be the same as what you and I might end up doing. That's another element of literacy." — Urva Shivatal (Kushi Baby), on designing for actual user behavior

"The problem is not in creating the plan. The problem is in getting people to manage the plan." — Dr. Anurag Aggarwal, on palliative care and implementation vs. diagnosis

Speakers & Organizations Mentioned

Panelists

Dr. Aperna Hay — Founder, Arman; Gynecologist & social entrepreneur (maternal and child health focus)
Urva Shivatal — Country Lead, Evidence & Impact, Kushi Baby; formerly Jhpiego
Dr. Anurag Aggarwal — Dean of Biosciences & Health Research, Ashoka University; Physician-scientist, genomics leader
Dr. Mitun Kumar Mitra — Professor of Physics, IIT Mumbai; Theoretical physicist working on living systems and TB analytics

Organizations & Initiatives

Arman — NGO providing mHealth solutions and training to community health workers across 7 Indian states; 500,000+ healthcare workers reached
Kushi Baby — Not-for-profit strengthening public health systems through data; works with Asha workers in Rajasthan and Maharashtra
Microsoft Research — Technology partner for ASHA chatbot; provides CSR/philanthropic support
Hopkins (Johns Hopkins University) — Collaborator on TB post-treatment outcomes modeling
National Disease Modeling Consortium — Set up by Ministry of Health & Family Welfare; brings together programmatic expertise, ICMR, epidemiologists
Palliium India — Partner in palliative care knowledge repository work
Bhashini — Multilingual language support platform referenced
Gemini — Language model platform referenced
Microsoft Gates Foundation — Supporting voice data benchmarking initiative
Government of India (MoHFW) — Created ANM protocols (2019 partial care guidelines); involved in health data exchange planning
Indian states: Telangana, Uttar Pradesh, Maharashtra, Rajasthan, Bihar, Odisha (mentioned as geographies of deployment or disease burden)

Technical Concepts & Resources

AI/ML Approaches

Retrieval-Augmented Generation (RAG): Core architecture used in both Arman and Kushi Baby chatbots to ground responses in validated protocols/guidelines
Large Language Models (LLMs): GPT-based and other models tuned on local data
Multimodal Input: Voice and text-based queries; voice-to-text conversion
Multi-Agentic Workflows: AI systems that cross-check and validate before responding
Massively Online AB Testing (PMAN) — Proposed framework for continuous, real-world validation of AI solutions in diverse user populations rather than centralized RCTs

Data & Evidence Methods

Epidemiological Modeling: Traditional statistical + machine learning models to account for heterogeneity
RCT (Randomized Controlled Trials): Acknowledged as insufficient for heterogeneous Indian contexts
Voice-to-Digital Data Conversion: Technology to capture interactions and build data repositories from voice-based interactions (Bhashini, others)
Digital Health Data Exchange: Proposed federated architecture allowing data use without centralized storage (DPDP Act–compliant)

Implementation Platforms & Tools

WhatsApp-based Chatbots: Leveraging existing user behavior; used by both Arman (ANMs) and Kushi Baby (ASHAs)
Voice-based Phone Interfaces: Call-in systems with auto-language recognition (referenced as in-test mode)
Learning Management Systems (LMS): App-agnostic platforms delivering video, simulations, interactive quizzes on protocols
RCH (Reproductive and Child Health) Database: Government database in states like Telangana; used to track outcomes and layer additional questions
Government Protocols: ANM protocols (35 high-risk conditions), color-coded, algorithmic, state-specific modifications

Key Datasets & Data Sources

Programmatic Data: Government TB, LF, and other disease surveillance databases
Local Surveys & Studies: Community-level data to supplement national datasets
Entomological Data: Largely absent for vector-borne diseases (malaria, dengue, lymphatic filariasis, encephalitis)
Socioeconomic & Cultural Data: Missing or underrepresented; critical for understanding heterogeneity

Health Domains Discussed

Maternal & Child Health: High-risk pregnancy detection, ANM decision support
Tuberculosis: Post-treatment outcomes, recurrent TB detection, screening in high-risk populations
Lymphatic Filariasis ("Elephantitis"): Neglected tropical disease modeling; vector biology; mass drug administration compliance
Anemia Detection: Community-level screening via digital tools
Palliative Care: End-of-life planning communication and adherence

Terminology & Concepts

Last-Mile Populations: Poorest, most remote, hardest-to-reach communities
Health Heterogeneity: Geographic, socioeconomic, linguistic, cultural variation requiring localized solutions
Data Solidarity: Legislative + educational approach to protecting data subjects while enabling research use
Data Representation: Ensuring underrepresented populations are included in training datasets
Data as Performance Metric vs. Opportunity: Reframing honest reporting of gaps/failures as diagnostic opportunity rather than accountability threat
Human-in-the-Loop: Hybrid AI-human decision-making, especially for safety-critical scenarios
Offline-Capable Tools: Solutions functional without reliable internet
Digital Literacy & Readiness: Ability to use smartphone/feature phone; language proficiency; familiarity with app/voice interfaces

Summary Table: Key Implementations

Initiative	Organization	Target Users	Technology	Coverage	Key Innovation
Integrated High-Risk Pregnancy Management	Arman	ANMs, medical officers, specialists	LLM chatbot (WhatsApp) + RCH database	7 states, 14,000 ANMs trained (scale to 12,000 by March)	Protocol-grounded RAG; state-specific modifications; 10-second latency; voice + text input
ASHA Sahili	Kushi Baby	ASHA workers (frontline community health)	LLM chatbot (WhatsApp) + government guidelines	Rajasthan, Maharashtra; 11,000 ASHAs	RAG-based safety guardrails; addressed counseling, nutrition, vaccination queries; confidence booster
TB Post-Treatment Outcomes	IIT Mumbai + Hopkins	District TB officers, state programs	Epidemiological + ML models	India-wide (programmatic data)	Local heterogeneity modeling; identifies lost-to-follow-up, migrant workers; tailored by geography
Lymphatic Filariasis Elimination	National Disease Modeling Consortium	State health programs, district officers	Statistical + ML modeling	Endemic districts (Bihar, Odisha, Maharashtra)	Integrates human behavioral data + entomological gaps; identifies "never treated" populations
Palliative Care Planning	Ashoka University (with Palliium India)	End-of-life patients & families	Multi-agentic GenAI workflow; voice interface	In development/test mode	Addresses implementation gaps rather than diagnosis; builds knowledge repositories; structured "I don't know" responses
Voice-Based Health Interface	Ashoka University (with Bhashini, Gemini)	Low-digital-literacy users	Phone call interface; auto-language recognition; voice-to-text	In test mode	Solves digital divide; no smartphone/app literacy required; multilingual (Hindi, English, Marathi pilot)

Critical Gaps & Future Work

Data Representation: Systematic inclusion of underrepresented populations (migrant mothers, tribal, religious minorities, adolescents)
Procurement Standards: Benchmarking framework for comparing AI health solutions across contexts
Health Data Exchange: Federated data architecture (DPDP-compliant) allowing research use without centralization
Entomological Data: Large-scale vector surveillance for TB, malaria, dengue, LF
Voice-to-Digital Infrastructure: Scaling technology to reliably convert voice interactions into structured data
Evidence for Scale: Understanding how solutions proven at 500–11,000 users perform at state or national scale

This summary preserves the transcript's emphasis on implementation realism, context-driven design, and the centrality of human infrastructure and data justice to any successful AI health deployment in India.