There is a statistic that should alarm anyone involved in healthcare technology: for every hour a clinician spends with a patient, they spend nearly two hours on documentation and administrative tasks in the EHR. This is not a technology problem that was accidentally created. It is a technology problem that was deliberately designed into the system — one click-heavy, checkbox-laden screen at a time.
The Electronic Health Record was supposed to be the digital transformation of medicine. Instead, it became the single largest contributor to clinician burnout. A 2024 Mayo Clinic study found that EHR-related stress is the strongest predictor of physician burnout, surpassing workload, specialty, and organizational factors. Emergency physicians spend an average of 4,000 clicks per shift. Primary care physicians spend 5.9 hours per day in the EHR, of which only 33% is during face-to-face patient encounters.
The industry’s response has been to bolt AI features onto these fundamentally broken systems. A chatbot here. A voice transcription widget there. A “smart” suggestion that requires three clicks to accept and five to dismiss. These additions do not solve the problem. They add complexity to an already overcomplicated system.
At IntelMedica, we are building something different: an AI-native healthcare operating system that treats intelligence as a foundational design principle, not an add-on feature. We call it the Intelligent UI.
Why Bolt-On AI Fails
To understand why the Intelligent UI needs to exist, you need to understand why the current approach — adding AI capabilities to existing EHR platforms — is structurally doomed to underperform.
The data model is wrong. Legacy EHRs were designed around billing, not clinical reasoning. The fundamental data structures — encounter-based documentation, problem lists organized by ICD codes, medication lists detached from clinical context — reflect the needs of the revenue cycle, not the clinical workflow. When you bolt an AI assistant onto this data model, the AI inherits all of its limitations. It can help you code a visit faster, but it cannot help you think about a patient differently.
The interaction model is wrong. EHRs are built on a request-response paradigm inherited from 1990s client-server architecture. The clinician asks for something (clicks a button, opens a tab, runs a search), and the system responds. AI-native design inverts this: the system anticipates what the clinician needs and surfaces it proactively. You cannot achieve this inversion by adding a sidebar to an existing application. It requires rethinking the entire interaction model.
The integration model is wrong. Existing EHRs are monolithic systems that treat external integrations as second-class citizens. FHIR APIs exist, but they are often read-only, rate-limited, and incomplete. Building an AI layer that needs to read patient data, write back recommendations, trigger workflows, and coordinate with external systems is fighting the architecture at every step.
The trust model is wrong. In a bolt-on approach, AI suggestions appear as foreign elements in a familiar interface. Clinicians learn to ignore them the same way they learned to ignore pop-up alerts — because the system has trained them that most automated suggestions are noise. An AI-native system can build trust through consistent, contextually appropriate behavior from the first interaction.
The Intelligent UI Concept
The Intelligent UI is not another EHR. We are not trying to replace Epic, Cerner (now Oracle Health), or Athena. These systems are deeply embedded in healthcare operations, and replacing them is a multi-year, multi-million-dollar undertaking that most organizations are not willing to attempt.
Instead, the Intelligent UI is a clinical operating system that sits alongside the EHR, integrated via FHIR and other standard interfaces, providing an AI-native workspace where clinicians actually do their thinking, documentation, and decision-making.
Think of it this way: the EHR remains the system of record. The Intelligent UI becomes the system of work.
This is not a novel concept in other industries. Software engineers do not write code in their version control system — they use an IDE that integrates with Git. Financial analysts do not build models in their accounting system — they use specialized tools that pull data from the general ledger. The Intelligent UI provides healthcare with the equivalent: a purpose-built workspace optimized for clinical cognitive work, backed by the EHR as the data layer.
Core Modules
The Intelligent UI is built as a modular platform. Organizations can deploy the modules they need without adopting the entire system. Each module is independently useful but becomes more powerful when integrated with others.
HIPAA-Compliant Messaging
Clinical communication today is fragmented across pagers, personal phones, EHR in-baskets, and ad-hoc messaging apps that may or may not be HIPAA-compliant. The Intelligent UI’s messaging module provides:
- End-to-end encrypted messaging with role-based access control.
- Context-aware threads that automatically link conversations to patients, encounters, or tasks.
- AI-powered message triage that identifies urgent communications and routes them appropriately.
- Automatic de-identification for messages that need to be shared outside the clinical team.
- Integration with EHR in-basket for messages that need to become part of the medical record.
The key differentiator is context. When a nurse messages a physician about a patient, the system automatically surfaces the relevant patient information — recent vitals, active medications, pending orders — so both parties have the same context without manual chart review.
Smart Scheduling
Scheduling in healthcare is a constraint satisfaction problem of staggering complexity: patient preferences, provider availability, room and equipment requirements, insurance restrictions, urgency levels, follow-up timing, and travel time for multi-appointment visits. Current scheduling systems handle a fraction of these constraints, leaving staff to manage the rest manually.
The Intelligent UI’s scheduling module uses constraint-based optimization to:
- Suggest optimal appointment times based on all relevant constraints.
- Predict no-shows using historical patterns and proactively fill gaps.
- Coordinate multi-appointment visits (lab draw, imaging, specialist consult) into efficient sequences.
- Identify scheduling conflicts and suggest resolutions.
- Adapt dynamically when appointments run long or providers fall behind.
Clinical Notes Assistance
This is the module that addresses the documentation burden most directly. Rather than a dictation tool or a template filler, the notes assistance module understands the clinical encounter and generates documentation that reflects it.
The system works in three modes:
Ambient mode: With the clinician’s and patient’s consent, the system listens to the clinical encounter and generates a draft note in real time. The draft is structured according to the organization’s preferred format (SOAP, H&P, specialty-specific templates) and includes relevant clinical codes. The clinician reviews, edits, and signs — turning a 15-minute documentation task into a 2-minute review.
Assisted mode: The clinician dictates or types key findings, and the system expands them into a complete note. “Patient presents with 3 days of productive cough, low-grade fever, CXR shows right lower lobe infiltrate” becomes a fully formatted note with the history, exam, assessment, and plan sections populated based on clinical context and the patient’s chart.
Review mode: The system reviews a completed note for documentation gaps, coding opportunities, and clinical inconsistencies. Did the note document the medication reconciliation? Is the assessment supported by the documented findings? Are there additional diagnosis codes supported by the documentation?
Clinical Decision Support
The Intelligent UI reimagines clinical decision support (CDS) as a continuous, contextual layer rather than a series of interruptive alerts.
Traditional CDS fires alerts when specific triggers are met — drug-drug interaction, allergy conflict, missing order. These alerts are right about 5% of the time and wrong about 95% of the time, which is why clinicians override more than 90% of them. The signal is buried in noise.
Our approach is different:
- Contextual surfacing: Instead of alerts, relevant clinical information is surfaced in the clinician’s workspace at the right time. When a clinician is reviewing a diabetic patient’s chart, the system surfaces the latest HbA1c, time since last retinal exam, and renal function trends — not as alerts, but as ambient context in the interface.
- Proactive recommendations: When clinical guidelines suggest a specific action (screening, referral, medication adjustment), the recommendation appears as an actionable item in the clinician’s task list, with the supporting evidence linked.
- Diagnostic reasoning support: For complex cases, the system provides differential diagnosis suggestions based on the documented findings, ordered by probability, with the evidence for and against each diagnosis explicitly stated. This is not the system making a diagnosis — it is the system organizing the clinical reasoning.
Training and Knowledge
Healthcare evolves constantly. New guidelines, new drug approvals, new evidence. Keeping clinicians current is a challenge that CME (Continuing Medical Education) addresses poorly — annual conferences and journal articles do not translate into real-time clinical practice.
The Intelligent UI’s training module provides:
- Point-of-care learning linked to current patient cases. When a clinician encounters an unfamiliar condition, they can access curated, evidence-based information without leaving the clinical workspace.
- Personalized knowledge recommendations based on the clinician’s case mix and identified knowledge gaps.
- Peer learning through anonymized case discussions within the organization.
- Compliance training integrated into the clinical workflow rather than siloed in a separate LMS.
EMR Integration Strategy
The Intelligent UI connects to existing EHR systems through a layered integration strategy. This is not optional — it is the core architectural decision that makes the system viable.
Layer 1: FHIR R4. The foundational integration layer uses HL7 FHIR R4 APIs, which are now mandated for certified EHRs under the ONC Cures Act. This provides read access to patient demographics, problems, medications, allergies, lab results, vital signs, encounters, and clinical notes. For Epic systems, we use Epic on FHIR. For Oracle Health, we use their Millennium FHIR APIs.
Layer 2: SMART on FHIR. For deeper integration, we use the SMART on FHIR framework to launch the Intelligent UI within the EHR’s application frame. This provides single sign-on, automatic patient context sharing, and the ability to write data back to the EHR through FHIR write operations.
Layer 3: CDS Hooks. The HL7 CDS Hooks specification allows the Intelligent UI’s clinical decision support to integrate directly into the EHR’s workflow. When a clinician opens a patient chart, prescribes a medication, or signs an order, the EHR can query the Intelligent UI for contextual recommendations that appear within the EHR interface.
Layer 4: Custom APIs. For EHR-specific functionality not covered by FHIR (scheduling, messaging, order entry), we build custom integration adapters. These are isolated modules that can be developed and maintained independently for each EHR platform.
Layer 5: Bulk FHIR. For analytics, population health, and training our clinical AI models, we use FHIR Bulk Data Access to export de-identified datasets from the EHR.
This layered approach means the Intelligent UI can provide value at any level of integration. An organization that only enables Layer 1 (read-only FHIR) still gets the documentation assistance and clinical decision support modules. Full integration unlocks the complete feature set.
Design Philosophy: Information Density Without Cognitive Overload
Healthcare UIs face a unique design challenge. Clinicians need access to enormous amounts of information — a patient’s medical history, current medications, recent lab results, imaging, notes from other providers, insurance details, scheduling constraints. Hiding information behind clicks and tabs forces clinicians to hold mental models of data they cannot see. Showing everything at once creates visual noise that obscures the signal.
The Intelligent UI’s design philosophy resolves this tension through three principles:
Progressive disclosure with AI prioritization. The system determines what information is most relevant to the current task and context, and surfaces it prominently. Less relevant information is available but de-emphasized. This is not a static layout — it adapts in real time based on the patient, the encounter type, and the clinician’s current activity.
Spatial consistency. Critical elements (patient identity, active alerts, navigation) maintain fixed positions. Contextual information flows into consistent zones. Clinicians build spatial memory of the interface, reducing the cognitive cost of finding information.
Minimal interaction cost. Every action in the Intelligent UI is designed to require the minimum number of interactions. If the system can infer the clinician’s intent from context, it pre-fills the appropriate values. If an action requires confirmation, it is a single click, not a multi-step wizard. We measure interaction cost in clicks-per-task and actively optimize it.
The aesthetic is intentionally restrained. No decorative elements. No gamification. No animations that do not serve a functional purpose. Healthcare UI design should respect the clinician’s attention as the scarce resource it is.
Technical Architecture
The Intelligent UI is built on a modern, high-performance stack designed for real-time clinical workflows:
Frontend: SvelteKit 2 with Svelte 5. We chose Svelte over React for a specific reason: performance. Clinical UIs cannot afford the rendering overhead and hydration costs of virtual DOM frameworks. Svelte compiles to vanilla JavaScript that manipulates the DOM directly, resulting in faster load times, smaller bundle sizes, and smoother interactions. Svelte 5’s runes system provides fine-grained reactivity without the complexity of React’s hook rules or state management libraries.
SvelteKit’s server-side rendering ensures fast initial page loads, while its client-side navigation provides the responsiveness of a single-page application after initial load. For offline scenarios (clinicians in areas with poor connectivity), we use service workers for background data synchronization.
Styling: TailwindCSS 4. Utility-first CSS ensures consistent visual design across the application without the specificity wars and dead CSS that plague traditional CSS architectures. TailwindCSS 4’s container queries enable responsive components that adapt to their container size, which is critical for a modular layout where panels can be resized.
Backend: FastAPI (Python). FastAPI provides the async-first, type-safe backend that clinical AI workloads demand. Python’s dominance in the ML/AI ecosystem means our clinical models, NLP pipelines, and decision support algorithms run natively without language boundary overhead. FastAPI’s automatic OpenAPI documentation and Pydantic validation provide the type safety and documentation that healthcare software requires.
Database: PostgreSQL. PostgreSQL is the backbone of our data layer. We use pgvector for embedding storage (clinical note similarity, semantic search), pg_trgm for fuzzy text matching (medication names, diagnostic terms), and PostgreSQL’s native JSONB for flexible clinical data structures that do not fit neatly into relational schemas. Row-level security policies enforce access controls at the database level.
Real-time: WebSockets. Clinical workflows are inherently real-time. When a lab result comes in, the clinician should see it immediately — not on the next page refresh. WebSocket connections provide bidirectional real-time communication for live updates, collaborative editing (shared care plans), and instant messaging. We use a pub/sub pattern where clients subscribe to channels (patient updates, team messages, system alerts) and receive events as they occur.
AI Inference: vLLM. Clinical AI models (note generation, decision support, NLP) run on self-hosted infrastructure using vLLM for efficient LLM serving. This keeps all patient data on-premise, eliminates per-token API costs, and provides the low latency that real-time clinical workflows require.
The Role of Ambient AI
The most powerful AI is the AI you do not notice. Ambient AI in the Intelligent UI operates continuously in the background, reducing friction without demanding attention.
Predictive navigation: The system learns clinician workflow patterns and pre-loads the next likely screen or dataset. When a clinician finishes a patient encounter, the system already has the next patient’s chart loaded. When a clinician opens the medication module, the pharmacy benefits are already queried.
Auto-population: Forms, orders, and referrals are pre-filled based on the clinical context. A referral to cardiology for a patient with a new murmur automatically includes the relevant exam findings, recent ECG results, and the referring clinician’s suspected diagnosis.
Smart defaults: Order sets and preference lists adapt based on the patient’s diagnosis, the clinician’s historical patterns, and current evidence-based guidelines. A clinician ordering antibiotics for community-acquired pneumonia sees the guideline-recommended options first, adjusted for the patient’s allergies and renal function.
Anomaly detection: The system continuously monitors incoming data for clinically significant changes — a lab result that is trending in a concerning direction, a vital sign that deviates from the patient’s baseline, a medication interaction that was not present before the most recent prescription. These are surfaced as subtle visual indicators, not modal alerts, proportional to their clinical significance.
Context switching: When a clinician is interrupted (a phone call, an urgent patient, a colleague’s question), the system preserves the exact state of their workflow. When they return, they resume exactly where they left off, with a brief summary of what they were doing and any updates that occurred during the interruption.
Comparison with Existing Solutions
It is worth acknowledging what the incumbents do well:
Epic has the deepest integration with the U.S. healthcare ecosystem and the most comprehensive feature set. Their recent AI investments (partnership with Microsoft, ambient listening via Nuance integration) are significant. However, Epic’s architecture is fundamentally monolithic, and their AI features operate within the constraints of an interface designed in the early 2000s. Customization requires Certified Epic consultants and months of build time.
Oracle Health (Cerner) has the technical advantage of a more modern architecture (cloud-native, FHIR-first) and Oracle’s infrastructure. Their generative AI features are promising but early. The Oracle acquisition has introduced uncertainty that has slowed adoption among some health systems.
Athenahealth has the most modern user interface among major EHRs and a strong cloud-native architecture. Their marketplace approach to integrations is appealing. However, their AI capabilities are limited compared to Epic and Oracle Health, and their market presence in large health systems is smaller.
The Intelligent UI is not competing with these systems. It is complementing them. We integrate with all of them. Our value proposition is that we can deliver AI-native clinical workflows faster and more effectively than the EHR vendors can retrofit their existing platforms. We are unburdened by decades of legacy architecture, regulatory certification dependencies, and installed-base compatibility requirements.
The incumbents will eventually build much of what the Intelligent UI provides. Our advantage is time. Healthcare organizations need relief from documentation burden and cognitive overload now, not in three to five years when the EHR vendors’ AI roadmaps mature.
What We Are Building Toward
The Intelligent UI is not the end state. It is the foundation for a healthcare AI ecosystem where clinical intelligence is a platform capability, not a feature.
In the near term, the Intelligent UI reduces documentation burden and surfaces better information at the point of care. In the medium term, it becomes the orchestration layer for autonomous agents that handle administrative workflows. In the long term, it provides the interface through which AI and clinicians collaborate on clinical reasoning itself.
Each step requires earning trust through demonstrated reliability, transparency, and measurable clinical value. We are not promising a revolution. We are building one, incrementally, with the rigor that healthcare demands.
IntelMedica is building the Intelligent UI for healthcare organizations ready to move beyond the EHR status quo. If your clinicians are drowning in clicks and your AI pilot projects are not delivering, we should talk. Visit intelmedica.com to learn more.