How Google's MedGemma is Helping AIIMS With AI Clinical Assistants

Outpatient departments (OPDs) in government hospitals across India are notorious for overcrowded waiting rooms, long queues, and overburdened doctors. Patients often spend hours waiting for consultations that may last only a few minutes, leaving clinicians little time to review medical histories, analyse reports, and make informed decisions.

To address this challenge, the All India Institute of Medical Sciences (AIIMS) is building two AI-powered clinical applications using MedGemma, Google's open-source AI model designed for medical text and image analysis, to help doctors spend less time on paperwork and more time with patients.

One of the applications, Aarogyam, generates patient summaries and notes before consultations to streamline outpatient workflows.

"Instead of spending time on symptoms and conditions, doctors can focus on the interaction and the care itself," Rajroshan Sawhney, Frontier & Applied AI Partnerships Lead (APAC), Google Research and DeepMind, tells AIM.

The second application, IndusDerma, is being developed as an AI-powered dermatology assistant that supports community health workers in identifying skin conditions while assisting dermatologists with clinical recommendations.

"The clinician always has oversight and authority. The AI acts as a clinical assistant to help augment them and save time so they can care for more patients," he explains.

Dr. Somesh Gupta, Professor of Dermatology and Venereology at AIIMS, tells AIM that he believes that MedGemma's multimodal capabilities could significantly improve clinical workflows.

"As a clinician working at AIIMS, where OPDs see a heavy patient load and remain extremely busy almost every day, what excites me about MedGemma is its multimodal capability—its ability to interpret text, medical images, and clinical context together," he states.

He believes that once trained and validated on Indian datasets, MedGemma could become an effective clinical decision-support system. However, he emphasises that AI should remain an assistant rather than a decision-maker.

"AI can support doctors by summarising records, identifying lesions and patterns, flagging relevant information, and reducing repetitive workload, but clinical decisions need human judgement, compassion, empathy, human touch, and accountability," Dr Gupta mentions.

Why Google Chose an Open Model

Unlike proprietary healthcare AI services, MedGemma 1.5, released earlier this year in January, is freely downloadable and can run entirely within an institution's own infrastructure.

"When you download the model, you control the infrastructure where it sits. The model comes to you—not the other way around," Sawhney explains.

According to Google, this approach offers three major advantages. Hospitals retain complete control over patient data, deployments remain compliant with local regulations, and institutions can fine-tune the models using their own datasets.

This is particularly important for dermatology.

"Global models often perform poorly on darker skin tones. AIIMS is customising MedGemma using Indian dermatology datasets so the model performs in the Indian context," he states.

Sawhney says MedGemma has now crossed 1 million downloads globally.

Beyond AIIMS, several healthcare organisations—including Apollo Hospitals, Fortis Healthcare, and Aravind Eye Hospital—are already using Google's diagnostic AI models for diabetic retinopathy, lung cancer, breast cancer, and chest X-ray screening.

Outside tertiary hospitals, Wadhwani AI is using the open-source model to develop HealthVaani, supporting Anganwadi workers, while digital health non-profit Khushi Baby has built an active case-finding platform on Google's Open Health Stack that has already reached more than 35 million beneficiaries.

Google is also working with the Translational Artificial Intelligence for Networked Universal healthcare, an AI Centre of Excellence at IISc, to encourage Indian researchers to build healthcare AI tailored to India's disease profiles.

Although AIIMS' work on MedGemma represents one of Google's most visible healthcare collaborations in India, the company says its ambitions extend far beyond AI models. Its larger objective is to help build the digital foundations for AI-enabled healthcare at national scale.

Building India's Digital Health Infrastructure

One of the central objectives of the Ayushman Bharat Digital Mission (ABDM), a government initiative to build a robust digital health infrastructure, is to improve healthcare accessibility with the same emphasis traditionally placed on brick-and-mortar healthcare facilities. Achieving that vision requires more than digitising hospitals. It demands interoperable health records, standardised data, and technology platforms that enable innovation at scale.

This is where Google sees itself playing a long-term role.

Through its research team, Google has been working with the National Health Authority (NHA), the implementing agency for ABDM, to develop open-source tools that simplify integration with the mission's digital architecture. The broader objective is to standardise medical records, enable interoperability, and equip hospitals with open AI models that clinicians can safely adapt for Indian healthcare settings.

A major initiative was creating an open-source compliance wrapper that reduced application integration timelines with ABDM from several months to just a few weeks.

Google is also collaborating with NHA to convert millions of unstructured medical records into Fast Healthcare Interoperability Resources, the global standard for healthcare data exchange. "Once you have a base layer of standardisation on health data, you can truly unlock services and longitudinal benefits for patient care," Sawhney remarks.

The company has also mapped more than 400,000 healthcare facilities onto Google Maps, enabling citizens to locate nearby PM-JAY (Ayushman Bharat) empanelled hospitals.

However, healthcare analytics expert Karan Tejpal believes that hospitals may be overestimating what the MedGemma model itself can do.

"It helps to be clear about what MedGemma actually is, because that decides where it can and cannot help. It is not a finished diagnostic product, and it is not meant to act as a doctor on its own," Tejpal states.

Instead, he sees its biggest contribution in easing the administrative burden on clinicians.

"A tool adapted from MedGemma can pull a patient's history into a usable summary, draft a first version of a radiology or discharge report for the doctor to check and correct, extract the key numbers from a lab report, and help flag the patients who look like the highest risk so they are seen sooner.”

Data Security Problem

Tejpal believes safety depends less on the AI model itself than on how hospitals deploy it.

"Safe is not a property of the model. It is a property of how the whole thing is set up."

He argues hospitals need strong consent mechanisms, encryption, role-based access, audit logs, and cybersecurity—not simply better AI models.

One significant advantage of MedGemma, he notes, is that it can operate entirely within a hospital's infrastructure.

"Because it is an open model that can be downloaded and run on a hospital's own servers, the data does not have to be sent to an outside company or a third-party cloud."

Although India's Digital Personal Data Protection Act provides a stronger data security framework, Tejpal believes implementation remains the real challenge. He adds that AI deployments will require continuous monitoring, cybersecurity audits, bias testing, and clearly defined accountability.

According to Tejpal, AI's greatest value lies in organising India's enormous volume of unstructured clinical information.

Doctors' notes, prescriptions, discharge summaries and referral letters remain largely inaccessible because they exist as handwritten records, scanned documents, or free text. Models such as MedGemma can convert these into structured data while improving triage, disease surveillance, multilingual patient communication, telemedicine support, and population-level analytics.

Still, he cautions against rushing adoption. "The sensible path is phased and validated, not a leap," Tejpal believes. "Public healthcare cannot run on a black box."