If you ask ChatGPT, Gemini or Grok-3 why a person with high fever speaks gibberish or hallucinates, it is highly likely that you’d get a false, that too confidently stated, output from these artificial intelligence engines.

Earlier, hallucination meant an illusory perception — seeing, hearing or feeling something that doesn’t really exist.

But after the arrival of AI and large language models (LLMs), the definition of hallucination has changed. And if you seek information on hallucination in a health context, you’d have to rummage through the entire response and skip the parts where the AI engine ‘hallucinates’ about its own errors.

The answers get confusing, possibly even risky, if you ask LLMs questions like “What does it mean if my test results are ‘positive’?” Or “What does ‘acute’ mean?” Or “How can I provide ‘support’ after surgery?”

This is where domain-specific models, or small language models (SLMs), come in — providing doctors, paramedics, researchers and others the exact information they seek.

Muting hallucinations

As more LLMs emerge, the leading models such as Grok-3, Gemini, Meta and Deepseek-R1 are getting stronger and evolving into ‘reasoning’ models.

They have disrupted the economy, and even science — opening up an array of possibilities, such as cutting down the time taken to develop new drugs or finding new molecules. But the outputs are not flawless. In fact, all such general-purpose models come with a big negative — hallucination.

Like humans who hallucinate when unwell, LLMs, too, often throw up unrelated stuff in their answers, requiring us to pay enormous attention to ensure the hallucinations don’t creep into the final output. And while the arts can afford hallucinations, science, particularly healthcare, doesn’t have that luxury. SLMs offer a safe way out.

SLMs are trained on knowledge exclusive to a domain — say, cardiology or cancer. Or even broad-spectrum healthcare, minimising the scope for hallucination.

Advantage SLMs

India has SLMs like Qure.ai and Vaidya.ai. “Healthcare is a big topic in India — we have a large population, several lifestyle diseases and insufficient radiologists and other doctors. It is a problem that must be addressed with AI,” says Srikanth Velamakanni, Co-founder and Group Chief Executive of Fractal AI. One of the first IT companies that built LLM models in the country, Fractal operationalised Vaidya.ai, an AI-based chatbot designed to be a ‘healthcare companion’.

“Multimodal AI models specifically trained with medical data, validated by doctors, and approved, where possible, by the regulators can dramatically improve healthcare outcomes efficiently,” he says.

Domain-specific models also need less computational capacities (or fewer high-end chips) and financial resources.

The field of biomedical natural language processing (NLP) has evolved from an early rules-based system to advanced LLMs, and then domain-specific SLMs, with researchers trying to build a domain-specific query engine for medicine and healthcare.

The introduction of ‘bidirectional encoder representations from transformers’, or BERT, in 2018 marked a significant shift, paving the way for domain-specific versions like ClinicalBERT and BioBERT.

These models, pre-trained on vast medical and biomedical datasets, showcased improved performance in tasks such as phenotype extraction and domain-specific responses.

From 2022 on, generative LLMs like BioGPT and BioMedLM emerged, focusing on generating coherent medical text and summarising research. Further advancements, exemplified by Google’s Med-PaLM and Med-PaLM 2, demonstrated LLMs’ potential in complex clinical reasoning and diagnostic support.

In the current phase, there is a move towards specialised SLMs and multimodality. Models like Microsoft’s RadPhi-2/3 integrate vision-language capabilities for radiology, while innovations like ClinicalMamba explore efficient architectures for long clinical contexts.

Several pharma companies are developing their own SLMs to exclusively develop molecules that suit their product roadmaps.

Exclusive LLMs

Ankit Modi, Chief Product Officer, Qure.ai, says that the adoption of exclusive LLMs in healthcare is transforming how clinical knowledge is accessed, interpreted and applied. Unlike conventional AI systems that are task-specific, such as interpreting medical images, LLMs possess the ability to comprehend and generate human-like interactions and dialogues.

“This linguistic intelligence enables real-time clinical decision support and seamless integration into complex healthcare workflows. Pairing LLMs with imaging models, such as a chest X-ray solution, can help generate simplified, patient-friendly reports and assist non-specialist clinicians in remote settings,” he says.

By connecting the dots, LLMs can suggest indications, and empower clinicians with faster, more informed starting points for decision-making. Beyond diagnostics, LLMs are reshaping administrative workflows, summarising complex clinical documentation and helping reduce the burden on care teams.

“What makes healthcare LLMs especially impactful is their potential to reduce clinician workload and democratise access to medical knowledge,” Modi says. However, he adds that successful implementation requires careful attention to data privacy, bias mitigation and responsible AI governance.

More Like This

Published on June 29, 2025