Healthcare

Overview

Healthcare emerged as the most substantively developed sector across the Summit, with AI applications spanning diagnostics, drug discovery, public health surveillance, frontline worker enablement, and administrative automation. India's combination of digital public infrastructure (ABDM, ABHA IDs), a massive and linguistically diverse patient population, acute specialist shortages, and a maturing startup ecosystem positions it as a genuine proving ground for healthcare AI that could serve the entire Global South. The central tension running through nearly every session is not technological capability—that is largely available—but the gap between functional pilots and population-scale deployment, driven by insufficient governance, uneven workforce readiness, and fragmented data standards. Speakers were unusually candid: India has built the foundation, squandered enough time on pilots, and now faces a 3–5 year window to demonstrate measurable population-level health impact before the demographic pressures of an aging population foreclose cheaper options 1. The stakes are not abstract—early sepsis prediction, silent heart attack diagnosis, breast cancer screening in women who have never accessed a radiologist, and TB detection at scale are the concrete applications being debated.

---

Key Insights

• Non-clinical workflows are the highest-ROI near-term opportunity. Administrative automation, drug manufacturing quality control, ambient clinical documentation, and population-scale screening programs have faster regulatory pathways, clearer cost-effectiveness, and proven demand compared to autonomous diagnostic AI 2. India's 84-site clinical trial network and AI-assisted pharma manufacturing are already generating 30–50% productivity gains in controlled settings 3.

• The ABDM architecture is necessary but not self-executing. India's Ayushman Bharat Digital Mission—86 crore ABHA IDs issued, interoperable federated infrastructure, privacy-first design—provides the enabling layer that most countries lack 4. But as Apollo's Sangita Reddy and others noted, infrastructure without incentive alignment, friction reduction, and workflow integration produces adoption rates far below potential 56.

• Federated learning resolves the privacy-scale tradeoff. Sending algorithms to data rather than aggregating patient records centrally—demonstrated in BODH's benchmarking approach and NHS COVID genomics work—allows Indian hospitals to collaborate on model validation without surrendering data sovereignty 78. This is the architecture that makes both DPDP compliance and population-scale AI possible simultaneously 4.

• Training AI on outcomes rather than physician annotations is methodologically critical. JPAL-partnered field studies on ECG interpretation showed that models trained to replicate clinical judgments automate existing errors; training against actual patient outcomes is more expensive but produces algorithms that genuinely improve upon current practice 9. This distinction is rarely enforced in procurement or validation standards.

• Voice and multilingual interfaces are equity requirements, not product features. Solutions that assume smartphone ownership, app literacy, or Hindi/English proficiency structurally exclude the populations with the highest disease burden. ASHA workers, the primary delivery channel for rural health, need voice-first, dialect-aware, offline-capable interfaces 1011. Bhashini's 22-language capability is a critical but insufficient foundation 12.

• AI's value in specialist-scarce India is democratization, not displacement. India has roughly 300 retinopathy of prematurity specialists for millions of preterm infants 13; AI extends that capability by orders of magnitude without requiring the specialists to move. The same logic applies to radiology for TB screening, ECG interpretation in primary care, and oncology decision support in district hospitals 91415.

• Neurosymbolic and explainable AI, not black-box generative models, are required for clinical adoption. Pharma and regulated healthcare contexts demand traceable reasoning aligned with medical guidelines. Clinicians reject systems they cannot interrogate, and regulators will not approve them 316. Explainability is simultaneously a safety requirement and a commercial differentiator in high-stakes sectors.

• India must develop and validate models on Indian populations. Algorithms trained predominantly on Western cohorts fail when applied to Indian patients—different disease prevalence, demographics, breast tissue density, genetic profiles, and clinical presentations 171819. Apollo's MDSAP/FDA validation partnerships and ICMR's consortium model represent the institutional scaffolding needed to close this gap 6.

• Governance, workforce training, and institutional readiness are the actual rate-limiting factors. Every session that moved beyond technology to deployment encountered the same bottleneck: health system staff who cannot troubleshoot devices, hospitals without data governance committees, state health programs without AI procurement frameworks, and clinicians who distrust systems they cannot audit 202122. The NamoShakti program—30,000+ breast cancer screenings in 4.5 months—succeeded precisely because it treated operational discipline as equal to technological innovation 18.

• Accountability frameworks have not kept pace with deployment reality. CDSCO regulatory approval does not resolve liability allocation between the AI developer, the hospital, and the clinician when an algorithm is wrong. This gap creates perverse incentives—either over-reliance on AI output or defensive non-use—and must be resolved before AI scales from diagnostic support to treatment recommendation 2324.

---

Recurring Themes

• "AI as workforce multiplier, not replacement" was the dominant clinical framing. Across cardiology 17, oncology 14, radiology 19, public health surveillance 25, and primary care 26, speakers independently converged on the same position: AI handles routine interpretation, flags urgent cases, reduces documentation burden, and extends specialist reach—but the clinician retains decision authority and accountability. The formulation from one session captures consensus: "A doctor using AI may replace a doctor without AI" 5.

• Pilots must become programs, or India wastes its window. "Pilotitis"—the proliferation of well-designed, well-funded demonstrations that never institutionalize—was named as a structural failure across at least six sessions 27215282930. The proposed remedies converge: outcome-linked financing, Health Technology Assessment (HTA) as a procurement gateway, multi-year government integration rather than philanthropic cycle funding, and measuring early detection rates and out-of-pocket cost reductions rather than algorithm accuracy.

• Local validation is non-negotiable and non-transferable. A model validated in Boston, Berlin, or even Mumbai does not automatically perform in Bangalore, Bhopal, or rural Bihar 31. Speakers from ICMR, WHO, Apollo, and multiple startups independently insisted that local re-validation against local patient populations, workflows, and infrastructure is mandatory—not a formality. This applies equally to models built abroad and to Indian models scaled across states with different epidemiological profiles.

• Data governance precedes data use. Whether discussing genomics 8, public health surveillance 25, drug discovery 32, or community health worker tools 33, speakers consistently identified the absence of clear data sovereignty frameworks, interoperability standards, and institutional governance committees as the primary obstacle—not algorithmic capability or compute access. DPDP compliance is necessary but incomplete; federated architectures, data trusts, and consent mechanisms require additional policy specification.

• Frontline worker trust and tool usability determine whether AI reaches the last mile. ASHA workers and ANMs are the actual delivery mechanism for AI-powered health interventions at population scale. Multiple sessions documented that surveillance-linked tools, administratively burdensome interfaces, and inadequate training destroy adoption—and that co-designing with workers before building, not after, is the only reliable path to uptake 3334. Digital literacy—including basic device troubleshooting—is as critical as clinical knowledge of AI applications 26.

---

Open Challenges & Tensions

• How do you generate evidence fast enough to be policy-relevant? Randomized controlled trials remain the regulatory gold standard for healthcare AI, but standard RCT timelines (3+ years to publication) are structurally incompatible with AI development cycles (6–18 months to meaningful version change) 9. Adaptive trial designs, pragmatic trials, and real-world evidence frameworks are proposed but not yet standardized in India. ICMR's clinical trial network provides infrastructure; the methodological consensus on how to handle model drift during a trial does not yet exist 17.

• Who bears liability when AI is wrong? CDSCO approval, ISO 42001 certification, and DPDP compliance all create accountability structures—but none clearly resolves who is liable when an AI-assisted diagnosis causes patient harm: the developer, the hospital, the clinician, or the regulator 2324. This gap is not theoretical; it is already shaping clinical behavior in ways that slow adoption and concentrate risk on frontline workers least equipped to bear it.

• Scale versus equity: does reaching 100 million patients mean excluding the most vulnerable? The most commercially viable healthcare AI applications—urban hospital efficiency, corporate wellness platforms, insurance risk stratification—tend to serve populations that already have healthcare access. The populations with the highest unmet need (rural, low-literacy, feature-phone users, tribal communities) require more expensive, lower-margin solutions 3536. No session resolved how to structure incentives so that the equity-critical use cases attract sustained investment.

• The validation bottleneck is institutional, not scientific. BODH's federated benchmarking model, ICMR's consortium, and Apollo's MDSAP partnerships all represent genuine progress—but India still lacks a systematized, adequately funded national infrastructure for validating AI on Indian clinical data at the pace the market is moving 67. The gap between the number of AI tools being deployed and the number being rigorously evaluated is growing, creating patient safety risk that current frameworks cannot detect.

• Sovereign AI in healthcare: build local or integrate global? Several sessions argued that India must develop small language models trained on ABDM-enabled longitudinal records, reflecting regional diversity 1, while others pointed out that no single actor—government or private—can build the full stack and that strategic partnerships with global platforms are necessary 372. The tension between data sovereignty requirements and the practical need to leverage frontier model capability remains unresolved, particularly for drug discovery and genomics applications that require access to global datasets 328.

---

Notable Examples

• NamoShakti breast cancer screening program scaled from concept to 30,000+ screenings in 4.5 months using thermal imaging (95% sensitivity, non-invasive, no radiation), mobile clinic deployment, survivor-led advocacy, and integrated digital referral pathways—demonstrating that operational discipline and multi-stakeholder coordination can achieve rapid health impact when political will is present 18.

• Apollo Hospitals' sepsis prediction system provides 24–48 hours of early warning across its network of 45 million registered users, with demonstrated reductions in ICU mortality. Apollo's EASE framework (Ethics, Adoption, Suitability, Explainability) and its MDSAP/FDA validation partnerships are being proposed as a replicable governance template for Indian healthcare AI 6.

• BODH (Benchmarking Open Data Platform for Health AI) federated benchmarking sends algorithms to hospital data rather than centralizing patient records, addressing both privacy compliance and India's fragmented data landscape simultaneously. This architecture allows institutions across geographically and epidemiologically diverse regions to participate in model validation without surrendering data control 7.

• CATCH Grant Awards 2026 for cancer AI, administered through an ICMR-led consortium, established standardized evaluation criteria for oncology AI tools and distributed grants for clinical decision-support systems specifically designed as workforce multipliers for India's specialist shortage—explicitly modeling the NCG (National Cancer Grid) collaborative approach over individual institutional competition 14.

• Karya's community data ownership model: 140,000+ workers performing 65 million+ AI annotation tasks across 70+ Indian languages, with dataset ownership retained by the communities generating the data. This model directly addresses the extractive data dynamic that has characterized Global South participation in AI development, and provides a template for how health data collected by ASHA workers could be governed 38.