In modern medicine, there is a fundamental contradiction that experts call the "Data Paradox." On one hand, Artificial Intelligence has the potential to save thousands of lives by learning from real patient histories and clinical cases. On the other hand, this data – Protected Health Information (PHI) – is the most sensitive and strictly guarded information in the world.
Because of vital regulations like HIPAA and GDPR, most of today’s medical AI models are trained on "sterile" data: textbooks, research papers, or synthetic datasets. They know the theory perfectly, but they lack access to the reality of clinical practice.
Today, we are sharing a case study on a convergence of technologies from Google Research, NVIDIA, Super Protocol, and Yma Health that solves this paradox. By fine-tuning MedGemma 27B within a Verifiable Confidential Computing environment, we have demonstrated that it is possible to bridge the gap between theory and practice – safely training on the most sensitive patient data to achieve a 9.4/10 clinician recommendation score.
For decades, the healthcare industry has sat on mountains of data that could revolutionize chronic disease diagnosis and therapy. However, the risk of data exposure during processing has remained an insurmountable barrier for the entire ecosystem – from agile HealthTech startups to established clinical providers.
Historically, stakeholders were forced to choose between two restrictive paths, neither of which fully solved the "Data Paradox":
The industry reached a stalemate: the data is too sensitive for the traditional cloud, but the models are too large for localized infrastructure. To break this bottleneck, Super Protocol shifted the paradigm from subjective trust to hardware-enforced, verifiable proof of the entire computation environment.
In healthcare, privacy must be verifiable, not just promised. While Trusted Execution Environments (TEEs) provide hardware-encrypted isolation for data in use, the TEE alone remains a "black box" – it secures the computation but cannot guarantee the provenance of the code or the integrity of the data pipeline.
Today, the primary barrier to adoption is the TEE orchestration challenge. Despite the availability of advanced TEEs, most clinical and development teams lack the specialized low-level expertise required to even initiate Remote Attestation. Without the ability to cryptographically prove that execution is occurring on genuine, untampered TEE (such as AMD SEV-SNP or Intel TDX), the trust chain is broken before it even begins.
When deploying large-scale models like MedGemma 27B, this complexity scales further, requiring a unified verification of the entire stack:
Establishing this trust chain manually is a major hurdle. This "verification gap" – the inability to initiate and bridge low-level hardware proofs with high-level AI workloads – means the environment remains unproven. Without such verification, processing PHI relies on subjective trust rather than objective proof, effectively halting the project's viability for clinical use.
To bridge the verification gap, Super Protocol provides a decentralized confidential compute layer that abstracts the complexity of TEE hardware into a universal, ready-to-use infrastructure. It removes the need for manual setup, providing a provider-agnostic environment where security is enforced by the protocol’s architecture rather than a central authority.
By automating the low-level handshake between hardware and software, Super Protocol establishes a seamless Trust Chain that extends from the physical silicon (NVIDIA/AMD/Intel) to the specific AI model and clinical data. This ensures that:
The collaboration with Super Protocol allowed Yma Health to bypass the TEE orchestration barrier by automating the entire execution lifecycle. By providing what effectively functions as Confidential DevOps-as-a-Service, Super Protocol managed the automated provisioning of TEE-enabled hardware and secure environments.
This transformed the complex MedGemma 27B pipeline into a secure, production-ready solution, allowing the AI team to treat highly sensitive, confidential infrastructure as easily as a public cloud and focus entirely on clinical fine-tuning:
To address the structural limitations of "sterile" datasets, Yma Health fine-tuned MedGemma 27B using real, protected clinical dialogues.
To align MedGemma with real-world clinical reasoning, Yma Health implemented a specialized training stack where the methodology was optimized for medical accuracy and the engine for high-performance execution.
The curation process prioritized clinical relevance and evidence quality over sheer volume, focusing on a high-signal 120,000-record dataset. To solve the "Data Paradox," the training centered on Metabolic Health, specifically GLP-1 receptor agonist (GLP-1RA) therapies and related cardiometabolic disorders.
The domain was structured into two subdomains:
This focused approach demonstrates Yma’s repeatable methodology: the same framework can be applied to any medical domain.
The training dataset was organized across multiple functional layers:
Confidential fine-tuning alone is not enough for medical AI. For a model to be useful in practice, clinicians must be able to run inference on real patient data under the same privacy guarantees as during training.
The core proof of this project’s success is the high rating given by medical professionals. More than ten independent practicing endocrinologists from UAE hospital networks validated the adapted MedGemma 27B model using a 100-question set in both English and Arabic.
In blind evaluations comparing Yma’s MedGemma mode against ChatGPT and human doctors, clinicians used a 5-point Likert scale to assess performance.
Image 1: Stacked barplots showing Safety and Conciseness benchmarks
The validation sessions demonstrated that the model identifies life-threatening complications that general-purpose models often miss.
Image 2: Smartphone screens showing Ozempic/Mounjaro side effect dialogues
Instruction tuning made LLMs responsive. Сonfidential inference made them respect user privacy. Confidential fine-tuning makes them stable and practical while caring about overall data security. Just imagine: you can replace tons of verbose instructions with only a few compact adapters that focus the model on what matters most. Thanks to open-source enthusiasts worldwide – and with Super Protocol’s confidential computing infrastructure, this is what we’ve just achieved!
– Daniil Pimanov, Head of AI, Yma Health
This project demonstrates a new paradigm of trust where Super’s verifiable Zero-Trust architecture and confidential computing transform sensitive medical data into life-saving innovations. It provides a clear framework for how foundation models like MedGemma 27B can reach their true potential and be safely applied to real-world clinical cases.
By shifting the foundation of security from contracts to architecture, we have shown that hospitals and AI developers no longer need to export sensitive data or rely on trust-based controls. This is the definitive path for bringing powerful AI into healthcare practice, while maintaining absolute privacy and regulatory compliance.
