nodae_bridge 1.0.5

Nodae BYOM SDK — NodaeBridge and NodaeModelAdapter for sponsor containers

nodae_bridge

Nodae BYOM SDK v1.0.4 — the bridge between your ML model and the Nodae federated healthcare platform by Intheris Health.

Overview

nodae_bridge is the Python package your Docker container uses when participating in federated training or inference across hospital sites. It provides:

• NodaeBridge — file-based I/O helpers (/data → /output)
• NodaeModelAdapter — the interface your model class must implement
• run.py — the container entrypoint dispatcher (python -m nodae_bridge.run)

Core principle: algorithms travel, data stays local. Patient data never leaves the hospital; your container receives a privacy-safe DataFrame and returns only model weights or aggregate statistics.

Installation

pip install nodae_bridge

Minimum Python version: 3.10. Add it alongside your ML framework in your Dockerfile:

RUN pip install nodae_bridge scikit-learn   # or torch, xgboost, etc.

Quick Start

from nodae_bridge import NodaeModelAdapter, ModelSpec, TrainingConfig, LocalTrainResult, InferenceResult
import pandas as pd

class MyModel(NodaeModelAdapter):

    def get_model_spec(self) -> ModelSpec:
        return ModelSpec(
            model_id="my-model-v1",
            name="My Federated Model",
            version="1.0.0",
            input_columns=["age", "length_of_stay", "drg_weight"],
            output_names=["risk_score"],
            min_samples=50,
        )

    def initialize(self, config: dict) -> None:
        pass

    def get_weights(self) -> bytes:
        return b""

    def set_weights(self, weights: bytes) -> None:
        pass

    def local_train(self, data: pd.DataFrame, round_number: int, config: TrainingConfig) -> LocalTrainResult:
        return LocalTrainResult(weights=self.get_weights(), sample_count=len(data), loss=0.0)

    def local_inference(self, data: pd.DataFrame) -> InferenceResult:
        return InferenceResult(prediction_distribution={"low": len(data)}, sample_count=len(data))

Your Dockerfile:

FROM python:3.11-slim
WORKDIR /app
RUN pip install nodae_bridge torch --index-url https://download.pytorch.org/whl/cpu
COPY model.py .
CMD ["python", "-m", "nodae_bridge.run"]

Data Schema

Your container receives local patient data at /data/input.parquet as a privacy-safe DataFrame. No PII is ever exposed.

Column	Type	Description
patient_id	str	Anonymized SHA-256 hash
age	int	Computed from DOB
gender	str	MALE / FEMALE / OTHER / UNKNOWN
admission_type	str	EMERGENCY / ELECTIVE / TRANSFER / AMBULATORY
discharge_disposition	str	HOME / TRANSFER / DECEASED / REHABILITATION
length_of_stay	int	Hospitalization days
primary_diagnosis	str	ICD-10 code
num_secondary_diagnoses	int	Count
drg_code	str	SwissDRG code
drg_weight	float	DRG cost weight
num_procedures	int	Count
num_medications	int	Count
num_lab_results	int	Count
insurance_type	str	Swiss insurance category

Always call fillna() before passing data to your model — not all columns are populated for every patient.

API Reference

Core I/O

Method	Description
NodaeBridge.get_data()	Load /data/input.parquet as a DataFrame
NodaeBridge.get_weights()	Load global model weights; b"" on round 0
NodaeBridge.get_config()	Load training/inference config dict
NodaeBridge.get_site_metadata()	Convenience wrapper for config["site_metadata"]
NodaeBridge.save_weights(bytes)	Write updated weights to /output/weights.bin (required)
NodaeBridge.save_metrics(dict)	Write aggregate metrics to /output/metrics.json
NodaeBridge.log(str)	Print timestamped message to stdout

Extended scope (byom_data_scope="extended")

Register with byom_data_scope="extended" to receive additional raw data files. Returns empty DataFrame/dict when absent (standard scope), so imports are always safe.

Method	Returns	Content
NodaeBridge.get_labs()	DataFrame	Per-result lab data: patient_id, loinc_code, test_name, value, unit, flag, date, …
NodaeBridge.get_vitals()	DataFrame	Latest vitals per patient: bp_systolic, heart_rate, temperature, weight_kg, …
NodaeBridge.get_procedures()	DataFrame	Per-procedure: patient_id, code, description, date, source
NodaeBridge.get_signals()	dict	Study-specific template signals: {patient_id: {signal_key: value}}

Feature engineering is the container's responsibility — the platform exposes raw data only.

FHIR scope (byom_data_scope="fhir")

Available when registered with byom_data_scope="fhir". The host writes one FHIR R4 Bundle per cohort patient under /data/fhir/. The standard input.parquet is always present alongside the bundles.

Method	Returns	Description
NodaeBridge.list_fhir_patients()	list[str]	Patient IDs (anonymized hashes) with FHIR bundles; reads index.json
NodaeBridge.get_fhir_bundle(patient_id)	Optional[dict]	Full FHIR R4 Bundle dict for one patient, or None

Each bundle contains: Patient (age, gender, LOS, DRG — no PHI), Conditions (ICD-10 + SNOMED), Observations (LOINC-coded labs and vitals), Procedures (CHOP), MedicationStatements, AllergyIntolerances, a Basic resource for template_signals, and an ImagingStudy reference when DICOM data is present.

for pid in NodaeBridge.list_fhir_patients():
    bundle = NodaeBridge.get_fhir_bundle(pid)
    entries = {e["resource"]["resourceType"]: e["resource"] for e in bundle["entry"]}
    # access any field directly — no platform-imposed field selection

SCAFFOLD (aggregation_strategy="scaffold")

Used when the study uses SCAFFOLD to correct client drift in non-IID populations.

Method	Description
NodaeBridge.get_control_variates()	Load global control variate from /data/control_variates.bin; b"" when not active
NodaeBridge.save_control_variate_delta(bytes)	Write per-site delta to /output/control_variates.bin

config["scaffold_enabled"] is true when the host has sent a global control variate. If save_control_variate_delta() is not called, the aggregator falls back to FedAvg for this site (backward compatible).

Imaging (imaging_access=True)

Available when registered with imaging_access=True and the SA host has BYOM_DICOM_STORE_ENABLED=True. A read-only DICOM store is mounted at /data/imaging/.

Method	Returns	Description
NodaeBridge.get_imaging()	Optional[Path]	Path to /data/imaging/, or None if unavailable
NodaeBridge.list_imaging_series()	list[dict]	Manifest: patient_id, study_uid, modality, slice_count, path

import pydicom
from pathlib import Path

base = NodaeBridge.get_imaging()
for series in NodaeBridge.list_imaging_series():
    if series["modality"] != "CT":
        continue
    slices = sorted((base / series["path"]).glob("*.dcm"))
    datasets = [pydicom.dcmread(str(s)) for s in slices]
    # HU windowing, segmentation, feature extraction — container's responsibility

Changelog

1.0.5 (June 2026)

• Added list_fhir_patients(), get_fhir_bundle() — FHIR R4 bundle access for byom_data_scope="fhir"
• Added FHIR_PATH class constant

1.0.4 (June 2026)

• Added get_imaging(), list_imaging_series() — DICOM store access
• Updated get_site_metadata() docs to include imaging.* keys

1.0.3 (June 2026)

• Added get_control_variates(), save_control_variate_delta() — SCAFFOLD support

1.0.2 (June 2026)

• Added get_signals() — template signal access
• Added get_labs(), get_vitals(), get_procedures() — extended data scope

1.0.1 (June 2026)

• Added get_site_metadata() convenience wrapper

1.0.0 (June 2026)

• Initial release: NodaeBridge, NodaeModelAdapter, run.py entrypoint

License

Proprietary — Intheris Health. Contact [e-mail] for access. \x00