AI can transform healthcare – if we transform our data architecture

Artificial intelligence (AI) use in healthcare has moved from theoretical to transformative in just a few years. Already, AI is helping us reimagine every aspect of healthcare – from administrative burden to diagnostics and patient outcomes.

At SingHealth in Singapore, for example, the Gen AI documentation system, Note Buddy, transcribes and summarizes patient visits into clinical notes across four languages simultaneously. At Mayo Clinic in the US, researchers have developed an AI tool called StateViewer, which helps clinicians more quickly and accurately identify brain activity patterns linked to nine types of dementia using a single, widely available scan. Meanwhile, predictive tools like CHARTWatch at UHN-Toronto, Canada, have reduced unanticipated mortality by 26% on the internal medicine ward.

These examples, among many others, illustrate the remarkable power of AI to improve human health and are why many believe AI transformation for global healthcare is inevitable.

That inevitability is, however, hitting a wall: healthcare’s long and complex history of data, spanning decades of disparate systems, incompatible formats, siloed records and legacy infrastructure. Outdated data structures constrain innovation by limiting AI to narrow, task-specific tools instead of enabling solutions that can reason, learn and act by accessing a full spectrum of multimodal healthcare data. The steps toward better, more personalized and accessible healthcare requires a bold reimagining of healthcare’s underlying data architecture.

Today, even the most digitally advanced organizations and nations lack the AI-ready data architecture needed to support the next generation of agentic and reasoning AI systems. For nation states to deliver on their commitment to embrace sovereign health AI as a national resource at a broad scale, within a tailored cultural, historical and ethical framework, they will need to reconceive their health data architecture to do so.

To create contemporary data architecture, the following concepts are critical:

1. A unified data pipeline

Most existing health data structures were built to provide batched updates while relying on manual inputs from the electronic health record, labs or imaging. Yet, increasingly, health data is and will be driven by real-time sensors and automated data inputs.

As a result, a modern data structure needs to take advantage of an intelligent, unified data pipeline that cleans, standardizes and labels incoming data in real time and converts this data directly into numerical representations intelligible to AI. This pipeline moves data away from fairly rigid relational databases and toward multi-dimensional graph databases that enable both real-time meaning and key context.

2. Making data ready, intelligible and navigable for AI

Even once processed, many critical pieces of health data elude traditional keyword searches and miss the hidden connections that can improve care and drive novel discovery.

There are newly established ways to extract maximal value from data that need to be embraced by healthcare, including vector databases, graph-based and hierarchical RAG (retrieval-augmented generation) foundations, extreme classifiers and world models. For example, the use of standard, searchable and vectorized data representations will ensure data is stored and annotated and used continuously in real time.

Vectorization, and similar methods, transform multimodal data (text, images, signals, genomic sequences, and others) into multidimensional numerical embeddings that capture relationships and meaning among data points. This allows AI systems to understand data across context, for example, recognizing similar clinical cases, relating symptoms to prior outcomes or retrieving relevant research findings with far greater precision.

One particular challenge is in clinical documentation – the vital insights, notes and narratives that make up a patient’s medical history. This data has proven especially difficult for AI to navigate while also being essential for its broad adoption. These advanced approaches are essential for such complex data and they make the data usable for AI insight generation, enabling fast retrieval and the real-time decision support that is essential to future clinical and research tools.

3. Unifying raw, curated and enriched data into a single, managed environment

Much healthcare data is siloed and fragmented across many different containers. Moving data into a single container can be costly and difficult to search. Health systems will need a data lakehouse as a foundation for storage. This enables secure access for analytics, AI and operations through standardized exposure layers that ensure both flexibility and trust. Data from electronic health records, laboratories, imaging devices, wearables and even patient apps can flow into a single interoperable environment as it is generated, becoming easily accessible for analytics and AI across one platform.

4. Ensuring data consistency, security and access across users and uses

The healthcare sector is relatively unique in facing the simultaneous complexity of multimodal data matched with the rigours of security regulations and ethics. While data is stored together across many organized containers (as in a lakehouse), it must also be highly secure and highly retrievable for AI to uncover healthcare insights.

An intelligent data fabric encompasses data scattered across many different containers and rules, and enforces data security and privacy. When data is requested by an AI application or clinician, it is first checked to validate that the requester has appropriate approvals and access controls, then it is pulled through the fabric, which knits the data together into a coherent and context-aware set of information for the end user.

5. Grounding data in trusted clinical and administrative insights

Health AI requires specific guardrails and context to ensure its insights align with operational and clinical guidelines. Knowledge graphs (a web of nodes, edges and labels comprising relationships among essential concepts and terminology) are one way to inform and ground AI decisions and outputs as context-driven and validated clinical insights. This helps make AI recommendations trustworthy and reviewable by care teams and operational counterparts.

By shifting the data model from a siloed, static retrieval system to an intelligent, continuously learning environment, every new data point can immediately inform care and drive research insights.

This new framework will also allow healthcare system AI to function in real time across all functions, where inputs in one system – for example, lab results or a vital sign monitor – are reflected throughout the network instantly via streaming updates. Furthermore, agentic AI tools built from such an architecture would be capable of automated reasoning, enabling much better and less costly healthcare at global scale.

From this architecture, AI agents will take advantage of real-time sensor data, including video or audio, to detect subtle signs that even a trained expert might miss. Just as importantly, these agents would learn from every outcome and feedback loop, improving their performance while remaining grounded in operational and clinical standards of care.

Other agents, trained for research, would simultaneously comb through the data for anomalies and patterns to answer investigative questions and drive breakthroughs, treatments and cures – learning continuously which data is valuable for each researcher and lab.

For patients, AI agents trained to answer health questions will stay up to date on the latest literature and can relay answers to basic care and patient education questions while triaging others to the respective clinician.

The vision for safe, patient-centric AI is possible today, but will only be realized by globally adopting a modernized data architecture. Through it, healthcare enterprises, governments and other health innovators can achieve our shared vision of transforming data into outcomes, and information into compassion, enabling us to reach more people in need.

Dr Matthew Callstrom, Mayo Clinic, and Dr Peter Lee, Microsoft, also contributed to this essay.