Building a vendor-agnostic patient health record search and summarization system with GEAP Agent Retrieval and Google ADK

By: Richard Mei & Tey Kitthajaroenchai

In modern healthcare environments, clinicians spend nearly two hours navigating electronic health record (EHR) systems for every one hour of direct patient care (Sinsky et al., Annals of Internal Medicine, 2016).Across health networks, a patient’s longitudinal medical history is rarely stored in one place—it is frequently fragmented across multiple disparate EHR vendors and ancillary clinical systems (e.g., inpatient vs. ambulatory EHRs, acquired specialty clinics, external laboratories, and imaging centers), as well as across hundreds of discrete FHIR resources (encounters, lab observations, medication orders, immunizations, conditions), unstructured clinical progress notes, and scanned attachments (e.g., faxed consultation reports, flowsheet PDFs, ECGs, clinical exam diagrams).

Traditional EHR search mechanisms fall short in clinical practice:

  • Keyword searches fail on clinical language: Querying “ASA” misses “aspirin”; querying “MI” misses “myocardial infarction”.
  • Structured filters miss data buried in narrative nuance: Progress notes, discharge summaries, and scanned documents remain unindexed dark data.
  • Clinicians lack longitudinal context: A single creatinine or HbA1c value is clinically meaningless without historical trajectory, reference ranges and surrounding context.
  • Black-box AI lacks clinical trust: Clinicians require verifiable source attribution, exact text citations, and visual pins on scanned documents before relying on an AI summary.

This article presents an architectural blueprint for an enterprise-grade Clinical Patient Health Record Search, Summarization, and Reasoning System. The architecture unifies fragmented clinical data across any healthcare organization (acute hospital networks, ambulatory care centers, academic health systems) and integrates cleanly across heterogeneous EHR ecosystems (Epic, Oracle Health/Cerner, MEDITECH, Athenahealth, or custom cloud FHIR repositories).

1. The core clinical drivers

Healthcare organizations worldwide share four fundamental clinical search and synthesis requirements:

  1. Overcoming semantic and acronym fragmentation: Clinical documentation relies heavily on shorthand, brand vs. generic drug names, and clinical synonyms. A search system must understand that “ASA”, “aspirin”, and “acetylsalicylic acid” represent the same therapeutic concept.
  2. Unified structured and unstructured retrieval: Providers must query structured diagnostic codes (LOINC, SNOMED CT, ICD-10, RxNorm) and unstructured narrative notes in a single query without manual ETL pipelines.
  3. Instant longitudinal trend visualization: Clinicians need immediate chronological trend lines complete with normal reference bands to spot acute changes (e.g., tracking renal decline or glycemic control across encounters).
  4. Agentic multi-step reasoning with strict attribution: Providers need an assistant capable of multi-step clinical deduction (e.g., “Review all cardiology notes and identify if ACE inhibitor dosage changed after the acute kidney injury episode”). Every claim must link directly to verifiable source records.

2. A unified vision: two solutions, one platform

By combining a low latency hybrid search and an agentic clinical assistant, the platform unifies two complementary paradigms into a single cohesive architecture allowing for clinicians to quickly analyze discrete structured and unstructured data along with conversational deep dives into the patient’s chart:

Figure 1: Unified Clinical Intelligence Platform combining Direct Analytical Search with an ADK Agentic Clinical Assistant.

  1. Direct search & longitudinal visualization: Designed for rapid clinical lookup, faceted exploration across 14 standardized clinical categories, query-level synonym expansion, and interactive physiological time-series charts with normal reference bounds.
  2. Interactive agentic clinical assistant: Powered by the Google Agent Development Kit (ADK) and deployed on GEAP Agent Runtime. Executes multi-turn clinical investigations, autonomous tool selection, multimodal image inspection with pinpoint coordinate overlays, and clinical findings timeline extraction.

3. Architecture

3.1. High-level conceptual architecture

At a high level, the system processes multi-format healthcare records through five streamlined stages using general-purpose GCP services:

Figure 2: End-to-end five-stage clinical data flow from EHR sources to the clinician web interface.

  • Stage 1 (Data Sources): Ingests structured HL7 FHIR R4 resources alongside binary scanned attachments (PDF, PNG, TIFF) from any provider’s EHR.
  • Stage 2 (Ingestion & Enrichment): Parses clinical codes (LOINC, SNOMED, ICD-10, RxNorm), normalizes clinical departments, and extracts text/metadata from scanned documents using Gemini multimodal models.
  • Stage 3 (Storage & Auto-Embedding): Stores raw unstructured documents in Cloud Storage (GCS) and indexes records in GEAP Agent Retrieval using native auto-embedding.
  • Stage 4 (Agentic Reasoning & Search): Coordinates hybrid retrieval, cross-encoder ML reranking, and autonomous multi-turn reasoning via the Google Agent Development Kit (ADK) on GEAP Agent Runtime .
  • Stage 5 (Provider Interface): Presents clinicians with a unified web portal combining faceted hybrid search, longitudinal lab trend charts, and an interactive “AI Mode” chat interface for conversational deep dives.

3.2. Detailed system architecture

To keep offline data preparation cleanly decoupled from low-latency clinical serving, the architecture separates into two operational planes: (A) an asynchronous ingestion and dual-layer indexing pipeline, and (B) an online hybrid retrieval and multi-agent reasoning tier.

Figure 3a: Offline Ingestion, Clinical Enrichment & Dual-Layer GEAP Agent Retrieval Indexing.

Figure 3b: Online Hybrid Retrieval, Cross-Encoder Reranking & Google ADK Agentic Reasoning.

4. Technical deep-dive: implementation & key patterns

4.1. Data modeling: dual-schema indexing & three-tier chunking in GEAP Agent Retrieval

Dedicated schemas for structured FHIR vs. unstructured documents

Healthcare data spans two fundamentally different modalities: compact, strongly typed HL7 FHIR resources (such as discrete lab results, medication orders, or immunizations) and variable-length unstructured clinical narratives (such as multi-page progress notes, operative reports, and OCR’d scanned attachments). Rather than forcing both modalities into a single monolithic table, the platform deploys two dedicated schemas in GEAP Agent Retrieval sharing strict patient and temporal isolation keys:

Schema Scope Key Schema Fields Clinical & Retrieval Purpose
Shared Isolation Keys patient_id, timestamp, date_display, document_category, department Scopes all hybrid searches strictly to a single patient, normalizes clinical specialty facets, and supports chronological ordering.
Structured FHIR Schema resource_type, resource_id, title, embedding_text, clinical_codes, code_displays, clinical_status Captures discrete FHIR R4 payloads (Observation, MedicationRequest, Condition, Immunization, Encounter) with standardized LOINC, SNOMED CT, ICD-10-CM, and RxNorm codes.
Unstructured Document Schema doc_id, parent_doc_id, chunk_tier, data_layer, chunk_text, summary, image_description, ocr_bounding_boxes, original_uri, file_type Supports three-tier adaptive chunking across clinical notes and scanned attachments, distinguishing document_summary parent vectors from granular chunk child vectors and linking to Cloud Storage URIs.

Auto-embedding configuration for structured (FHIR) data

Rather than maintaining external embedding pipelines prior to ingestion, both collections leverage GEAP Agent Retrieval’s native server-side auto-embedding capability. For structured FHIR resources, a dedicated multi-field text_template combines clinical taxonomy, specialty context, narrative summary, and standardized code descriptions—intentionally omitting unstructured visual fields so discrete clinical records remain tightly focused:

"text_template": "{resource_type} | {department} | {title}: {embedding_text} ({code_displays})"

Three-tier chunking & dual-layer strategy for unstructured data (notes & scanned docs)

Indexing unstructured clinical documents presents a classic granularity trade-off: indexing long encounters as monoliths dilutes semantic focus and exceeds embedding windows, whereas blindly splitting short forms destroys context. During ingestion, an adaptive three-tier chunking strategy processes unstructured notes and OCR’d attachments according to their extracted text length and content profile:

1. Tier 3 — Pointers & placeholders (`< 150` characters):

  • Handling: Known scanner artifacts, fax cover logos, or empty stubs are suppressed and never embedded. Clean, clinically meaningful short records are embedded intact as a single vector without chunking.

2. Tier 2 — Administrative & short clinical forms (`150` to `1,500` characters):

  • Handling: Embedded intact as a single vector, preserving complete context for short referral notes, vaccination records, and consent forms.

3. Tier 1 — Long clinical notes & multi-page attachments (`> 1,500` characters):

  • Handling: Produces a dual-layer parent-child representation (data_layer):

  • `document_summary` (Parent Vector): A single holistic summary vector capturing high-level clinical takeaways across the entire document (~75% smaller token footprint than full clinical notes), optimized for broad UI triage and first-pass agent exploration.

  • `chunk` (Child Vectors): Multiple granular child chunks split along clinical section headings, isolated tables, or sliding windows for high-precision semantic and lexical retrieval.

Ingestion and query-time parent resolution:

  • At ingestion: Tiers 2 and 3 produce single self-contained vectors, while Tier 1 indexes both granular child chunk vectors and a linked parent document_summary vector sharing parent_doc_id.

  • At query time: When hybrid search matches a Tier 1 child chunk, the retrieval layer resolves parent_doc_id to fetch the complete parent document text from the structured store / Cloud Storage before passing context to Gemini—combining pinpoint chunk recall with complete clinical context.

The 3-tier chunking and ingestion strategy addresses the unique structural and noise challenges of medical records by optimizing both retrieval accuracy and processing costs:

  • Effective noise suppression: Over 73% of scanned attachments are administrative noise or scanner driver logos (such as TWAIN branding). Filtering these out entirely (Tier 3) prevents irrelevant assets from polluting the vector index and wasting storage or embedding compute.
  • Intact preservation for short documents: Administrative forms, consent documents, and cover sheets (Tier 2) are kept intact as single vectors. This avoids fragmenting short texts that naturally fit within a single embedding window, ensuring complete semantic context is preserved.
  • Granular precision with contextual safety for long notes: Dense clinical narratives and multi-page records (Tier 1) are split into granular child chunks for precise retrieval matching, while retaining a parent summary vector. This hybrid approach ensures that a specific finding can be located via search while still feeding the complete parent document to Gemini during generation to eliminate clinical information loss.

4.2. Search pipeline: relevance gating, parallel Hyhrid search & ML reranking

Searching across 14 clinical categories simultaneously requires balancing semantic recall, exact code/lexical precision, and low-latency interactive responsiveness. In production, chaining a live generative LLM call onto the critical retrieval path introduces tail-latency variability; therefore, the search pipeline decouples generative synonym expansion asynchronously while executing category retrieval in parallel.

Figure 4: Parallel per-category hybrid search, non-blocking synonym expansion, and cross-encoder reranking workflow.

Step-by-step search workflow

  1. Unfiltered semantic pre-gate (phase 0): The backend executes a lightweight unfiltered vector probe across the patient’s chart to identify which of the 14 clinical categories actually contain candidate records relevant to the query—avoiding wasted fan-out requests against unrelated clinical domains.

  2. Parallel per-category hybrid search with RRF (phase 1): Across all candidate categories discovered by the pre-gate, the backend fires concurrent hybrid searches in parallel (rather than looping sequentially). Within each category, GEAP Agent Retrieval fuses dense semantic vector similarity and sparse BM25 lexical matching using Reciprocal Rank Fusion (RRF)—capturing conceptual matches via dense embeddings while anchoring exact drug names, dosages, and diagnostic codes via BM25.

  3. Cross-encoder ML reranking (phase 2): Because RRF scores are rank-position-based and cannot be compared directly across independent category queries, merged candidates are re-scored in a single batch call via the Discovery Engine Ranking API, producing calibrated relevance scores on a normalized 0.0 to 1.0 scale and rendering ranked results immediately to the clinician.

  4. Non-blocking asynchronous synonym expansion (parallel enrichment): Concurrently—and off the critical blocking path for initial result rendering—the backend checks an in-memory/Redis clinical acronym cache (covering high-frequency shorthand such as “ASA”, “CHF”, “MI”, or “HbA1c”) and falls back to an asynchronous Gemini Flash call for novel queries to expand clinical synonyms (e.g., “ASA” → “aspirin”, “acetylsalicylic acid”, “antiplatelet”). Because GEAP Agent Retrieval’s dense embeddings already handle semantic synonym recall during retrieval, these expanded terms stream cleanly to power deterministic snippet highlighting in the UI without delaying search results or invoking per-record LLM calls.

Deep technical pain point: Score Floor & Adaptive Elbow Cutoff

In clinical search, surfacing low-relevance noise erodes clinician trust and increases cognitive load. Rather than returning fixed top-K lists, the platform applies a two-stage pruning strategy over the calibrated cross-encoder scores:

  • Minimum confidence Score Floor: Discards baseline noise records that fall below a configurable relevance threshold, ensuring tangential matches are never displayed simply because a category contains data.

  • Adaptive Elbow Cutoff: Walks the remaining descending score curve and detects the sharpest relative drop between consecutive results—dynamically truncating the tail at the natural relevance boundary.

Sample code for Score Floor & Adaptive Elbow Cutoff. This example uses 0.05 as floor, which can be adjusted based on specific use cases.

def apply_calibrated_elbow_cutoff(
    ranked_results: list[dict],
    floor: float = 0.05,
    drop_threshold: float = 0.40,
) -> list[dict]:
    """
    Illustrative pruning logic over calibrated 0.0-1.0 cross-encoder scores.
    Production floor and drop_threshold parameters are calibrated per category
    via offline clinician-annotated golden sets and online interaction evals.
    """
    valid = [r for r in ranked_results if r.get("_score", 0.0) >= floor]
    if len(valid) <= 1:
        return valid

    scores = [r["_score"] for r in valid]
    max_drop, elbow_idx = 0.0, len(valid)

    for i in range(len(scores) - 1):
        if scores[i] > 0:
            pct_drop = (scores[i] - scores[i + 1]) / scores[i]
            if pct_drop > max_drop and pct_drop >= drop_threshold:
                max_drop = pct_drop
                elbow_idx = i + 1

    return valid[:elbow_idx]

Calibrating thresholds via offline & online evaluations:

Values such as floor = 0.05 and drop_threshold = 0.40 serve only as illustrative baselines. Because score distributions vary between terse structured lab records and long narrative consultation notes, production deployments tune thresholds systematically through:

  • Offline clinical evaluations: Benchmarking against clinician-labeled golden query-relevance datasets to jointly optimize precision@K and false-omission rate (recall)—ensuring clinically significant records are not overly pruned.

  • Online telemetry & safe UI fallbacks: Monitoring live click-through depth, query reformulation rates, and category expansions. For clinical safety, categories whose top candidate falls below the active threshold are collapsed into an expandable “Low-Confidence / Filtered Categories” drawer (rather than hidden irreversibly), keeping the primary workspace clean while preserving one-click clinician override visibility.

4.3. Longitudinal observation tracking & visual charting

An isolated lab value or vital sign rarely tells the full clinical story. For example, a serum creatinine of 1.4 mg/dL carries very different clinical urgency depending on whether it reflects a stable multi-year baseline in chronic kidney disease or an acute spike following a new medication order. Requiring clinicians to open dozens of individual lab reports across encounters to mentally reconstruct trajectories introduces heavy cognitive load and risks missing subtle inflections.

When search surfaces a quantitative observation with a standardized LOINC code (such as 2160-0 for Serum Creatinine), clinicians can launch an interactive longitudinal trend view in a single click—aggregating historical values, measurement units, and physiological reference bands into an immediate visual timeline that highlights out-of-range excursions at a glance:

Step-by-step implementation:

  1. LOINC-Driven Retrieval: When search returns an observation containing a recognized LOINC code (e.g., 2160-0 for Serum Creatinine), the system enables a “Track” action.
  2. Chronological Aggregation: The backend queries all historical records matching that LOINC code for the patient, extracting the numerical value (valueQuantity.value), measurement unit (valueQuantity.unit), and reference range limits (referenceRange[0].low and high).
  3. Time-Series Analysis: The server computes clinical indicators—minimum, maximum, latest value, and abnormal flag count—and returns a structured timeline.
  4. Interactive Canvas Rendering: The client renders an interactive HTML5 canvas line chart displaying historical data points, tooltips, and green shaded bands indicating the physiological reference range.

4.4. The agentic reasoning engine: Google ADK & GEAP Agent Runtime

For multi-step clinical questions that span encounters, medication changes, lab trajectories, and scanned specialty reports, the platform deploys an autonomous multi-agent architecture built with the Google Agent Development Kit (ADK) and hosted on GEAP Agent Runtime.

UX Entry point & seamless query handoff

Search and conversational reasoning share a cohesive, two-stage user experience rather than operating as disconnected silos:

  • Single-click escalation from search (shared seed query): Clinicians typically begin in the primary search bar with concise clinical terms (e.g., “ASA and renal impairment”). Clicking “Dive Deeper in AI Mode” from the search results view automatically hands off the active search query (alongside the bound patient_id session context) to seed the ADK agent’s opening turn—eliminating duplicate typing.

  • Dedicated multi-turn conversational input: Once inside AI Mode (or when opened directly for chart synthesis), clinicians interact via a dedicated multi-turn chat input optimized for compound clinical questions and follow-up dialogue (e.g., “Was the ACE inhibitor held during that November admission, and when was it restarted?”), streamed back in real time over Server-Sent Events (SSE).

Figure 5: Google ADK root orchestrator agent, polyglot tool routing, and selective multimodal sub-agent on GEAP Agent Runtime.

Key agent architecture patterns

  1. Session-scoped tenant isolation: A before_agent_callback intercepts incoming requests and binds the authenticated patient_id directly into the ADK session state. Crucially, LLM tool declarations omit `patient_id` as a model-controlled parameter; every tool reads ToolContext.state[“patient_id”] deterministically at execution time, making cross-patient data leakage via prompt injection or model hallucination structurally impossible.

  2. Polyglot patient-scoped tool suite: Vector databases excel at semantic and hybrid text discovery, whereas relational databases and FHIR servers excel at deterministic filtering, counts, and canonical ID lookups. Accordingly, the root ADK agent orchestrates tools across purpose-built backing stores—exposing structured EHR/database endpoints cleanly via Model Context Protocol (MCP) tools or REST adapters:

    • search_records (backed by GEAP Agent Retrieval): Executes filtered hybrid search across clinical narratives—starting with data_layer=“document_summary” for token-efficient broad exploration before drilling into granular chunk records—with automatic query deduplication.

    • filter_records (backed by Structured FHIR Store / AlloyDB via MCP): Performs deterministic metadata queries, chronological date-range filtering, and exact record counts (e.g., listing all active MedicationRequest orders or Immunization records).

    • get_record_detail (backed by Structured FHIR Store / AlloyDB via MCP): Retrieves the authoritative, complete FHIR resource or full parent clinical note text and standardized codes by unique record identifier.

    • get_patient_findings (backed by Pre-Indexed Findings Store via MCP): Queries the ontology-normalized clinical findings index (Section 4.5) to answer broad longitudinal problem, procedure, allergy, or medication history questions instantaneously.

  3. Two-phase multimodal strategy (offline ingestion vs. selective on-demand inspection): Running live multimodal LLM inference across scanned PDFs on every user query would introduce unacceptable latency and cost. Instead, visual document processing is split across two stages:

    • Ingestion stage (default — zero query-time vision aatency): During asynchronous batch ingestion, multimodal Gemini Batch / Document AI pipelines transcribe scanned text, synthesize structured image_description summaries, and index OCR text-block bounding boxes. As a result, standard search queries, text-span visual highlights on scanned PDFs, and over 90% of conversational agent turns resolve immediately against pre-indexed data with zero live vision model invocation.

    • Query stage (selective selegation via `AgentTool`): The root agent delegates to the specialized clinical_image_analyst sub-agent (invoking analyze_image against raw images in Cloud Storage without exposing gs:// URIs in prompts) only selectively during AI Mode turns when a clinician poses a targeted visual question about non-textual features—such as anatomical exam sketches (e.g., diabetic foot-exam diagrams), handwritten margin annotations, or complex multi-column nursing flowsheets—streaming live tool progress (“Inspecting scanned diagram…”) transparently in the chat UI.

Multimodal coordinate pointer grounding

When referencing scanned medical documents (such as faxed consultation notes, multi-column flowsheets, or clinical intake forms), standard LLMs often summarize page text without showing clinicians where a finding appears on the page. The platform addresses this by leveraging Gemini’s multimodal 2D spatial pointing capabilities—instructing the visual tool to return normalized (x, y) coordinate pointers with values from 0.0 to 1.0 (corresponding to horizontal left-to-right and vertical top-to-bottom offsets, which can be emitted directly via structured JSON schema or rescaled from Gemini’s native 0–1000 [y, x] spatial grid):

{
  "findings": [
    {
      "label": "Decreased breath sounds at lung base",
      "detail": "Physical exam finding noted in 2021-11-02 pulmonary consultation note",
      "pointers": [{"x": 0.42, "y": 0.68}]
    }
  ]
}

In the document viewer, the frontend renders interactive pin overlays positioned directly over the relevant passage or table cell using percentage offsets (left: 42%; top: 68%) so clinicians can visually verify scanned source evidence immediately:

4.5. Clinical findings extraction microservice

Unstructured progress notes frequently capture critical clinical context—such as historical symptoms, resolved episodes, or discontinued medications—that never make it into structured EHR problem lists.

Platform integration: connecting findings to aearch & the ADK Assistant

Rather than operating as an isolated standalone service, the Clinical Findings pipeline acts as an asynchronous clinical entity index persisted in the structured clinical store that enriches both interaction paradigms:

  • In the Direct Search Portal (Paradigm A): Powers a proactive “Patient Problem & Findings Timeline” view alongside faceted search, allowing clinicians to browse longitudinal conditions, procedures, symptoms, medications, and allergies extracted from free-text notes and cross-referenced against structured FHIR problem lists.

  • In the ADK Clinical Assistant (Paradigm B): Exposes pre-aggregated, ontology-coded findings via the get_patient_findings tool so the agent can answer broad synthesis prompts (e.g., “Summarize all active versus resolved cardiovascular conditions documented across notes”) in milliseconds without reading hundreds of raw clinical notes at query time.

Four-stage findings pipeline

  1. Entity Extraction & Offset Grounding: Extracts clinical mentions asynchronously using langextract and Gemini, mapping them directly to exact character offsets for verbatim source provenance.

  2. Classification, Negation Filtering & UI Routing: Classifies mentions by semantic class and status while filtering out negated terms, routing active and historical findings into six primary clinical UI categories.

  3. Ontology Normalization & Deduplication: Maps surface text to standard terminologies (SNOMED CT, ICD-10-CM, RxNorm) to merge duplicate concepts, compute timeline statistics, and reconcile against structured FHIR resources.

  4. Grounded Overviews via Provenance Lookup: Retrieves source passages using direct offset pointers to generate concise chronological summaries with inline citations, using hybrid search only for context expansion.

5. Enterprise advantages

Deploying this architecture delivers measurable clinical, operational, and engineering benefits:

  • Unified Healthcare Workflow: Combines high-speed analytical search and multi-turn conversational reasoning in a single workspace with one-click query escalation.

  • Data Sovereignty & HIPAA Compliance: Runs in Google Cloud VPC-SC under BAA without using PHI for model training.

  • Verifiable Attribution & Guardrails: Directly links claims to source data via text spans, citations, and visual pin overlays.

  • Optimized Latency & Cost Efficiency: Reduces LLM token load significantly using parent summaries and offline batch enrichment.

  • Universal EHR Interoperability: Built on open FHIR, standard medical terminologies, and MCP interfaces without vendor lock-in.

  • Instant Longitudinal Visibility: Visualizes time-series trends and clinical timelines instantly across historical records.

  • Fully Managed Operational Scale: Leverages serverless GEAP Agent Retrieval and Runtime for automatic scaling and maintenance.

6. Conclusion & the path forward

The future of clinical documentation systems lies in bridging the gap between high-speed analytical search and deep conversational synthesis. By combining Google Cloud’s GEAP Agent Retrieval with the Google Agent Development Kit (ADK), healthcare providers can break down data silos, extract insights from unstructured notes and scanned documents, and give clinicians their time back.

Actionable next steps for health tech teams:

  1. Standardize the FHIR pipeline: Establish automated extraction pipelines converting EHR resources and scanned document attachments into standardized FHIR R4 JSON/textproto.
  2. Deploy the Vector Search Collection: Implement the 25-field schema with native auto-embeddings and dual-layer storage in GEAP Agent Retrieval.
  3. Roll out the Direct Search Portal: Give clinicians immediate access to faceted hybrid search, query-level synonym highlighting, and longitudinal LOINC observation tracking.
  4. Introduce Agentic Reasoning Gradually: Deploy the ADK agent on GEAP Agent Runtime to assist with complex clinical summaries, starting with specialized sub-agents for medical document inspection.

By building on an open, vendor-agnostic foundation, healthcare engineering leaders can build a future-proof clinical intelligence platform that scales seamlessly across hospitals, specialties, and EHR ecosystems.

10 Likes

I am impressed, my son is Doc student in Paraguay, where he interns is a precarious hospital. Ideas like this is the future comes quickly. I’ll be grateful to share with him and help to develop whatever you need help. Thanks. AMAZING

2 Likes