A healthcare AI platform needed to ingest ECG recordings from every major device manufacturer and process them through a unified analysis pipeline. In theory, the DICOM standard should make vendor neutrality straightforward: every compliant device produces structured data that any compliant system can read. In practice, each manufacturer's DICOM implementation diverges enough to break that assumption. Trailing zeros in one vendor's amplitude values, renamed classification labels in another's firmware updates, and bitmap-encoded filter settings in a third's proprietary extensions meant that treating all DICOM files identically would silently corrupt the data clinicians rely on.

ML LABS built a vendor-neutral ingestion pipeline with per-manufacturer parsing adapters, measurement normalization that preserves clinical precision, deduplication at the storage layer, and a shared ECG library that provides unified access across all formats — including a research SDK for programmatic data retrieval.

Per-Vendor DICOM Parsing

Each manufacturer's DICOM output is treated as a distinct dialect. A shared ECG library provides the common data structures and normalization utilities, while vendor-specific adapters handle the format divergences that matter clinically.

graph TD
    A["Raw DICOM File"] --> B["Vendor Detection"]
    B --> C["Vendor A Adapter"]
    B --> D["Vendor B Adapter"]
    B --> E["Vendor C Adapter"]
    C --> F["Unified ECG Model"]
    D --> F
    E --> F
    F --> G["Validation &<br/>Storage"]

    style A fill:#1a1a2e,stroke:#e94560,color:#fff
    style B fill:#1a1a2e,stroke:#ffd700,color:#fff
    style C fill:#1a1a2e,stroke:#0f3460,color:#fff
    style D fill:#1a1a2e,stroke:#0f3460,color:#fff
    style E fill:#1a1a2e,stroke:#0f3460,color:#fff
    style F fill:#1a1a2e,stroke:#16c79a,color:#fff
    style G fill:#1a1a2e,stroke:#16c79a,color:#fff

Each incoming file is routed to the correct vendor-specific adapter automatically. Two different vendors' DICOM files containing the same clinical recording differ in measurement encoding, waveform byte ordering, and classification conventions. Running them through the same parsing logic produces values that look plausible but are clinically wrong — the kind of error that surfaces months later when a cardiologist notices a measurement discrepancy between recordings of the same patient on different devices.

Measurement Normalization

ECG measurements — intervals, amplitudes, axes, and derived scores — are the values clinicians read directly on reports and use for diagnostic decisions. The normalization layer solves three distinct problems across vendors.

First, measurement formatting differences: one vendor stores amplitude values as 1.200 while another stores the same value as 1.2. In a clinical context, trailing zeros can imply measurement precision. The pipeline normalizes all measurement values to a canonical format with explicit precision metadata, so downstream consumers can distinguish genuine precision from formatting artifacts.

Second, classification mapping: device manufacturers assign proprietary names to ECG classifications, and firmware updates can rename standard labels to product-specific ones. The pipeline maps all vendor-specific labels to a unified clinical meaning, so the same condition produces the same classification regardless of which device recorded it — without requiring redeployment when a vendor changes their labeling.

The hardest interoperability problems are not protocol-level. They are the subtle measurement normalization differences between vendors that produce clinically incorrect values if handled generically.

Filter Metadata Normalization

ECG signal filters affect waveform morphology and measurement accuracy. The filter settings recorded at acquisition time influence how AI models and clinicians interpret the waveform.

Each manufacturer represents filter settings in a different proprietary format. The pipeline normalizes all vendor-specific representations into a unified filter record: which filters were active, their cutoff frequencies when available, and whether the filter state was read from metadata or inferred from device defaults. This record travels with the waveform through every downstream step — AI inference, clinical display, and research export all consume the same normalized filter information.

Waveform Deduplication and Storage

A platform that ingests ECGs through multiple paths — web upload, SMB file share, EHR integration — will receive the same recording more than once. Without deduplication, the platform stores three copies and displays them as three separate studies.

ML LABS implemented deduplication at the storage layer based on clinically meaningful content rather than raw file bytes — the same recording arriving through different paths resolves to a single stored record. This kept storage costs predictable and eliminated a class of clinical workflow bugs where duplicate records confused clinicians reviewing a patient's ECG history. The immutable storage layer satisfies regulatory data integrity requirements — every access traces back to an immutable deduplicated record.

SDK and Research Access

Researchers need programmatic access to ECG waveform data across all ingested formats — loading recordings from two different manufacturers and comparing waveform characteristics should not require format-specific handling in research code.

ML LABS built SDK waveform loading that abstracts vendor format differences behind a unified interface. The SDK loads waveforms from any supported format, applies the same normalization pipeline used in clinical processing, and returns a common data structure with measurement values, filter metadata, and waveform samples. Research workflows can also access raw DICOM metadata alongside the normalized representation when studying format-specific artifacts.

When Prior Experience Matters

Vendor-neutral parsing is tractable when the format landscape is bounded — a known set of manufacturers with documented DICOM conformance statements. It becomes categorically harder when a new vendor's implementation deviates significantly from the standard or when firmware updates change encoding conventions without notice.

Teams that have navigated the undocumented behaviors of multiple ECG device manufacturers — trailing zeros, renamed classifications, bitmap filter fields, private DICOM tags — move through new vendor onboarding at a pace that first-time builders cannot match.

First Steps

  1. Catalog your device landscape. Inventory every ECG device manufacturer, model, and firmware version across your clinical sites — this defines the parsing scope and identifies the vendor-specific behaviors your pipeline must handle.
  2. Build one vendor adapter end-to-end. Parse a single vendor's DICOM output into a common ECG model with full measurement normalization and filter handling, then validate against clinician-reviewed reference values before adding a second vendor.
  3. Anchor regression tests to real device output. Collect anonymized DICOM files from every supported device and firmware version in your network, and test against these real files rather than synthetic data — synthetic files miss the encoding quirks that cause production failures.

Practical Solution Pattern

Build vendor-neutral ECG ingestion by treating each manufacturer's DICOM output as a distinct parsing dialect with a dedicated adapter, while all adapters converge on a single common model with deterministic measurement normalization, explicit filter metadata, and deduplication at the storage layer.

This architecture works because it isolates vendor-specific complexity at the adapter boundary. The clinical data model, storage layer, and every downstream consumer — AI inference, clinical reports, research exports — remain stable as new device manufacturers are added. The measure of success is whether a clinician can trust that the same patient's ECG produces identical measurements regardless of which device recorded it. If your organization needs to scope a vendor-neutral ingestion architecture, a Strategic Scoping Session can pressure-test the parsing requirements and normalization strategy before engineering work begins.

References

  1. NEMA. DICOM Standard. National Electrical Manufacturers Association, 2024.
  2. U.S. Food and Drug Administration. Cybersecurity in Medical Devices. Regulatory Reference, 2025.