Beyond large language models: Why the future of healthcare AI will depend on specialised, purpose-built models
By Ashissh Raichura, Founder and CEO, Scanbo Technologies
The most impressive healthcare AI demo right now is a model passing a medical licensing exam. The most useful healthcare AI right now is something far narrower, running in a clinic nobody writes about.
I have spent ten years putting diagnostic AI into real clinical settings. The distance between those two sentences is where most healthcare AI projects die.
A large language model reads text. Clinical care begins before any text exists. A patient walks into a clinic with chest pain and there is no document to read. There is a waveform, a blood pressure value, a glucose reading, and a doctor with very little time. Someone has to turn that into something a model can work with, and the quality of that step decides everything downstream.
This is the part the industry keeps skipping. We learned early that if you do not control the device, you do not control the signal, and if you do not control the signal, you cannot stand behind the output. A model trained on clean research recordings then meets mains interference, a loose electrode, a patient who will not sit still, and a health worker who received twenty minutes of training. The model is usually fine. The input is not.
Purpose-built begins with acquisition. Before we applied to the Central Drugs Standard Control Organisation in 2024, we screened almost 250,000 ECGs across geographies, age groups and ethnicities. Of those, 129,000 came from India. None of it was synthetic. A patient at a Tier-2 clinic in Gujarat does not look like the open-source ECG libraries from MIT and Stanford that most teams start with. Those libraries are a reasonable place to begin. They are not an Indian clinic.
It also means building a model that refuses to answer. Our system checks the quality of the recording and asks the health worker to take it again when it falls below the accepted threshold. A general-purpose model will answer anything you ask it. A clinical model has to recognise when the data in front of it is not good enough to read, and say so. That behaviour is harder to build than the interpretation itself, and it never appears in a demo.
And it means scope you can defend in writing. Our engine reads a single-lead ECG. That is one view of the heart, seen from the front. Stating the boundary is part of the product, because a clinician who does not know what the system cannot see has no way to use it safely.
Regulation makes all of this unavoidable. A clearance applies to a defined version of a model doing a defined task with a defined failure mode. There is no pathway for a system that does everything and changes every Tuesday. Teams building on frontier models that update without notice will discover this at submission, which is the most expensive moment to discover it.
The economics point the same way. A narrow model running on infrastructure you control has a cost per test a small clinic can recover. A general model called per query, priced by a vendor whose roadmap you do not influence, does not.
None of this argues against large language models. They are very good at what they are built for, including the documentation burden that eats a doctor’s day. The error is treating one architecture as the answer to every clinical problem.
India has an advantage here that it is not using. Our own healthcare system is among the hardest operating environments anywhere, and a model proven on Indian signal quality, Indian patient populations and Indian clinic conditions will hold up almost anywhere else. Building narrow and building local are the same decision.