Stop Funding Federated Learning in Indian Healthcare
Apollo Hospitals validated its AI cardiovascular risk tool using federated learning through Microsoft Azure, built on ten years of anonymized data from 400,000 patients. IIT Guwahati signed an MOU with Coventry University on AI and healthcare collaboration. AIIMS partnered with IIIT-Delhi for innovation in “mobile health, AI, interoperability.” IMARC Group published a report on India’s “nascent yet rapidly evolving” federated learning in healthcare market.
The announcements keep coming. The pilots keep launching. The conference panels keep convening.
And the district hospital in Nashik that’s supposed to be a node in this distributed intelligence network can’t maintain consistent patient records. Their billing runs on decade-old software. Their scheduling is an Excel sheet.
This is not an infrastructure gap. It’s a category error. And the capital flowing into federated learning pilots is capital not flowing into the problems that actually need solving.
The three missing prerequisites
Federated learning assumes three things. In Indian healthcare, none of them are true.
The data doesn’t exist in usable form. Federated learning says: keep data where it is, train locally. This assumes the data can train a model. The ABDM’s own integration data tells the story: out of 35 crore health records linked through the registered health facility registry as of February 2024, only 2% — roughly 7 lakh — were shared by private healthcare providers. Some hospitals have their EHRs as scanned PDFs, which aren’t even compatible with ABDM systems.
Walk into a tier-2 hospital: patient records split across paper files, partial EMR entries, WhatsApp messages between doctors, lab results in PDFs. Different diagnostic codes at each clinic. No shared patient ID.
You can’t federate what isn’t digital, structured, or standardized.
The compute doesn’t exist. Tier-2 and tier-3 cities still lack reliable internet access and trained IT teams — this is from ABDM’s own assessment of onboarding challenges. Each federated node needs to run model training locally and synchronize updates reliably. Most tier-2 hospitals I’ve visited don’t have a full-time IT person.
The incentive doesn’t exist. Indian hospitals are capacity-constrained, not quality-constrained. A tier-2 hospital doesn’t lose patients because its diagnostic AI is worse than the hospital across town. The market rewards beds and doctors, not algorithms. Private providers have already invested in their own digital infrastructure; additional investment to align with ABDM “does not have an adequate business case” — that’s the assessment from NATHealth’s 2024 report on catalyzing digital health in India.
The entropy problem
Federated learning is trying to decrease entropy across a system with no central ordering force.
In thermodynamics, you need external energy to create order in a closed system. Federated learning proposes creating useful models from distributed chaos without any central pipeline imposing structure.
A patient visits three clinics for the same chronic condition. Clinic A has a partial EMR entry. Clinic B has a handwritten note. Clinic C has a WhatsApp photo of a prescription. Federated learning would train a model across all three simultaneously. The model learns from garbage.
I’ve built systems that ingest unstructured content — PDFs, documents, free-text notes. The lesson is always the same: data cleaning and standardization is 70% of the effort. You solve this by controlling the pipeline end-to-end. That requires centralized ownership. The exact opposite of what federation proposes.
What actually works
Every successful Indian healthcare AI product has the same architecture. Every single one.
Tricog. Cardiac diagnostics for remote areas. ECGs captured on edge devices, sent to a central hub. AI plus remote cardiologists analyze centrally. Deployed in 14+ countries, served 6 million patients, 334 employees, ₹77.7 crore revenue as of March 2025. $30 million raised.
Tricog controls the data pipeline end to end. If they’d tried to federate across hundreds of hospitals, they’d still be negotiating data formats.
Niramai. Breast cancer detection via thermal imaging. Centralized model. Controlled data capture — same thermal sensing device, same protocols. A state-wide study funded by the Government of Punjab screened 15,069 women across 183 locations in 18 months. Detection rate of 0.18%, positive predictive value of biopsy at 81.81%. Now commercially available in 22 countries.
Works because they own the quality of the input, not because they federated across messy hospital data.
Qure.ai. Chest X-ray AI for TB. Cloud service. Hospitals send images, get results. Deployed in 2,600+ sites across 100+ countries. Five million AI-powered chest X-rays in partnership with AstraZeneca across 20+ countries. At Maha Kumbh 2025, their qXR flagged 36.22% of X-rays as abnormal, with 12% showing presumptive TB signs.
One vendor. Full pipeline control. One high-value clinical problem. Clean data. Centralized intelligence, distributed access.
| Federated Learning Thinking | What Actually Works |
|---|---|
| Train models where data lives | Control the data pipeline end-to-end |
| Preserve local autonomy | Impose standardization centrally |
| Distribute compute | Centralize intelligence, distribute access |
| Privacy-preserving by design | Privacy through controlled access |
The ABDM numbers don’t tell the full story
India has been building the interoperability layer for over a decade. The numbers sound impressive at first glance: 84.79 crore ABHA IDs created as of January 2026, 82.69 crore health records linked, 230,000+ health facilities registered.
But look closer. Only 2% of linked records come from private providers. Large hospitals struggle to integrate existing systems with ABDM. Data format compatibility is a mess — scanned PDFs masquerading as digital records.
And the incentive problem remains: hospitals that have already invested in their own digital infrastructure see no business case for additional ABDM alignment. The ones that haven’t invested don’t have the capacity to participate in federated learning even if the standards were perfect.
ABDM is solving the right problem — interoperability. But federated learning is jumping three steps ahead, assuming a level of digital maturity that doesn’t exist outside tier-1 hospitals.
What I got wrong
I initially thought the problem was technical — better privacy-preserving protocols, more efficient gradient compression, smarter aggregation algorithms. We explored federated approaches for a healthcare client early on. The technical problems were solvable.
The structural problems weren’t. You can’t engineer around missing incentives. You can’t compress gradients when there’s no gradient to compress because the local data is unusable. The failure mode wasn’t the federation protocol. It was everything that had to be true before the protocol even mattered.
The question worth asking
The real question isn’t how to make federated learning work in Indian healthcare. It’s why we’re trying to make it work when centralized alternatives are already proving effective.
Tricog, Niramai, and Qure.ai didn’t wait for perfect interoperability. They built controlled pipelines for specific clinical problems. They’re serving millions of patients. They’re generating revenue. They’re saving lives.
Federated learning research is intellectually interesting. The privacy guarantees are real. The distributed training efficiency gains are measurable in the right contexts.
But Indian healthcare is not the right context. Not yet. Maybe not for a decade.
The capital going into federated learning pilots would be better spent on the boring infrastructure work: standardizing EMR formats, training hospital IT staff, building incentive structures for data sharing, solving the actual interoperability problems that ABDM is tackling.
Or better yet: fund more companies like Tricog. Centralized intelligence. Distributed access. Controlled pipelines. Proven clinical value.
That’s the architecture that works. Everything else is a conference paper.