Skip to content
SETHIOS — Become Your Identity

RESEARCH / 001

Experimental

Privacy-Preserving Federated Learning for Clinical Decision Support

Collaborative healthcare AI without centralizing sensitive clinical data.

Type
Master research direction
Year
2026
Status
Experimental
Fields
Federated learning · Healthcare AI · Clinical decision support · Privacy
Research outline

ABSTRACT / RESEARCH SUMMARY

Sensitive clinical data is commonly distributed across institutions that cannot simply pool raw records. This research investigates federated learning as a collaborative model-development approach for clinical decision support, using a selected MIMIC-III experimental cohort and centralized, IID, non-IID, and privacy-aware scenarios.

The study treats performance, privacy parameters, distribution shift, and resource consumption as connected concerns. OpenHealthFL provides the engineering framework direction through which experiments, hospital participants, aggregation, privacy controls, and research artifacts can be represented.

RESEARCH / PROBLEM

Collaborative intelligence meets institutional and privacy boundaries.

Centralized learning assumes that clinical data can be brought into one environment. Healthcare data silos, governance requirements, infrastructure variation, and sensitivity make that assumption difficult and often undesirable.

Federated learning keeps selected training data within participating environments and exchanges model updates. Its usefulness still depends on statistical heterogeneity, privacy mechanisms, model behavior, orchestration, and the resources available to each participant.

OBJECTIVES

What the investigation evaluates.

Evaluate a privacy-preserving federated-learning approach for clinical decision support in distributed and resource-constrained healthcare settings.

  1. 01Establish centralized logistic-regression and MLP reference scenarios.
  2. 02Compare federated learning under IID and non-IID institutional partitions.
  3. 03Examine differential-privacy parameters and their privacy–utility implications.
  4. 04Observe the computational impact of privacy mechanisms in the represented experiment context.
  5. 05Translate the experimental architecture into the OpenHealthFL research framework.

METHODOLOGY / PROTOCOL

A staged experimental path from cohort to privacy analysis.

The protocol selects an experimental cohort, preprocesses available panel values, constructs centralized and hospital-partitioned scenarios, trains reference and federated models, evaluates predictive metrics, and observes privacy and resource implications.

FIGURE / 001
  1. 01

    Dataset

  2. 02

    Cohort selection

  3. 03

    Preprocessing

  4. 04

    Hospital partitioning

  5. 05

    Local training

  6. 06

    Aggregation

  7. 07

    Evaluation

  8. 08

    Privacy analysis

Experimental methodology from cohort selection through privacy analysis.Source: Research protocol

DATASET / COHORT

MIMIC-III experimental cohort.

Selected experimental cohort; these counts do not describe the full MIMIC-III dataset.

129
Initial admissions
128
Adult admissions
103
Admissions with panel data
81
Unique patients
32.04%
Positive class rate
ClassAdmissions
Class 070
Class 133

MODEL / ARCHITECTURE

Reference MLP and federated system.

FIGURE / 002
Multilayer perceptron architecture used as a centralized reference.Source: Experimental model definition
FIGURE / 003
Federated learning architecture. Raw clinical data remains at hospital nodes; model updates move through aggregation.Source: OpenHealthFL experimental framework

EXPERIMENTAL SCENARIOS

S0

Centralized logistic baseline

Logistic B reference model trained in a centralized experimental setting.

S1

Centralized MLP

Multilayer perceptron reference under centralized training.

S2

Federated IID

FedAvg with participating data represented under an IID partition.

S3

Federated non-IID

FedAvg under heterogeneous institutional distributions.

S4

Federated + privacy

Federated experiments with differential-privacy parameters.

57

Experiments in research UI

18 / 18

Resource experiments

60

Federated rounds

EXPERIMENTS / RESULTS

Model comparison.

Values are reported experimental results; bars use a 0–1 scale and labels remain the authoritative values.

ModelAUROCAUPRCF1
Logistic B0.71880.61790.6667
Centralized MLP0.58330.50850.4444
FedAvg IID0.54170.46240.1667
FedAvg Non-IID0.59380.4973—
Logistic B
0.7188
Centralized MLP
0.5833
FedAvg IID
0.5417
FedAvg Non-IID
0.5938

AUROC / SCALE 0—1

DIFFERENTIAL PRIVACY

Privacy parameters in experiment context.

At fixed δ, the represented higher noise multiplier corresponds to a lower ε. This illustrates a privacy–utility trade-off; it is not a production privacy guarantee.

Noise σEpsilon εDelta δ
410.08761e-5
84.64631e-5

RESOURCE ANALYSIS / R5 DP VS R4 CPU

Observed computational impact.

Comparison specific to the represented differential-privacy experiment and CPU baseline; not generalized to deployment.

+242%

CPU usage

+159%

Runtime

+1.07%

RAM

DISCUSSION

The experiment exposes trade-offs, not a deployment verdict.

Within the represented results, Logistic B provides the strongest listed AUROC, AUPRC, and F1 values. The federated scenarios remain sensitive to partitioning and experimental conditions; their value cannot be reduced to a single ranking.

Increasing the represented noise multiplier from σ = 4 to σ = 8 corresponds to a lower listed ε at the same δ, while privacy, utility, runtime, and resource behavior must be interpreted together. These experimental observations motivate inspectable architecture and resource-aware federation.

SCIENTIFIC CONTEXT

Limitations

  1. 01The selected experimental cohort is limited and is not the full MIMIC-III dataset.
  2. 02Institutional partitions are simulated rather than a real multi-hospital deployment.
  3. 03The test sample and represented model comparisons limit generalization.
  4. 04The work is experimental and does not constitute clinical validation.
  5. 05Federated performance is sensitive to non-IID distributions.
  6. 06Differential privacy introduces utility and computational trade-offs.

RESEARCH DIRECTION

Future work

  1. 01Evaluate broader and more diverse datasets.
  2. 02Pursue real multi-institution validation.
  3. 03Study secure aggregation and stronger privacy mechanisms.
  4. 04Refine resource-aware federated participation.
  5. 05Integrate interoperability and clinical exchange concerns.
  6. 06Extend reproducible OpenHealthFL research artifacts.