Article Navigation Heading

Personalized Clinical Decision Support in Multi-Hospital Acute and Chronic Care Using Privacy-Preserving Federated Transfer Learning


Ruchi Gupta1, Sandeep Gupta2*, Tadiwa Elisha Nyamasvisva3, Anupama Sharma4, Nitin Tyagi5 and Ashif Ali6

1Postdoctoral Researcher, Kuala Lumpur University of Science and Technology, Malaysia and Faculty of Department of Information Technology, Ajay Kumar Garg Engineering College, Ghaziabad, U.P., India

2Department of CSE, Sharda University, Greater Noida, U.P., India and International Visiting Professor, Faculty of Engineering , Science and Technology, Kuala Lumpur University of Science and Technology, Malaysia

3Faculty of Engineering, Science and Technology, Kuala Lumpur University of Science and Technology, Malaysia.

4Faculty of Department of Information Technology, Ajay Kumar Garg Engineering College, Ghaziabad, U.P., India.

5Postdoctoral Researcher, Kuala Lumpur University of Science and Technology, Malaysia and Department of CSE-AIML, Accurate Institute of Management and Technology, Greater Noida, U.P., India

6Department of CSE, Aravali College of Engineering and Management, Faridabad, Haryana , India

Corresponding Author E-mail: guptasandeep1093@gmail.com

Download this article as: 

ABSTRACT:

Machine learning-based clinical decision support systems (CDS) are becoming more common in healthcare, but their implementation in a multi-hospital setting is limited by data privacy laws, institutional heterogeneity, and personalization requirements. Centralized learning approaches cannot align with healthcare data governance stipulations, whereas current federated learning (FL) frameworks exhibit a lack of robust convergence and deteriorate in performance when applied to non-independent and non-homogeneously distributed (non-IID) clinical data. We address these issues with a personalized federated transfer learning (FTL) model for clinical decision support in which population-level representation learning is kept apart from hospital-specific adaptation. All sites share one global backbone, while each hospital keeps its own adapter layers and prediction head on premises. Training combines proximal-regularized federated optimization with secure aggregation and differential privacy using privacy budgets ε = 1, 5, and 10 and δ = 1 × 10⁻⁵. The proposed framework is tested on simulated multi-hospital partitions derived from two publicly available clinical datasets for two clinically distinct tasks: acute sepsis survival prediction and chronic kidney disease (CKD) classification. In the sepsis task, the proposed approach achieves an AUROC of 0.69 ± 0.01 (95% CI: 0.68–0.70; p < 0.001 vs. FedAvg), which is better than standard federated baselines of FedAvg (0.43 ± 0.09) and FedProx (0.48 ± 0.12), and the removal of hospital-level personalization decreases performance (AUROC = 0.53 ± 0.14). The framework achieves an average AUROC of 0.86 ± 0.04 on the CKD task, which is more accurate and less variable across inter-hospital settings than local training and global federated strategies. All findings suggest that explicit personalized federated transfer learning can support potentially deployable clinical inference systems in heterogeneous medical organizations without disclosing raw patient information. Key limitations include the use of virtual hospital partitions rather than real multi-institutional data, the absence of real-world latency and workflow evaluation, and the need for external validation before clinical deployment.

KEYWORDS:

Clinical decision support systems; Federated transfer learning; Privacy-preserving machine learning; Personalized healthcare; Sepsis dataset

Introduction

Although federated learning-based Clinical Decision Support (CDS) elicits increased research interest, there remains a gap in the literature to allow concomitant privacy preservation, non-independent and non-identically distributed (non-IID) clinical data resistance, and institution-level personalization in practical multi-hospital settings. The current literature on CDS tends to address one of these issues in isolation—decentralized optimization, transferability between tasks, or formal privacy assurances—but almost none typically combine all three into a deployable framework. Consequently, existing methods cannot make credible and clinically implementable predictions in heterogeneous institutions while meeting regulatory and operational demands (Naik et al.).33

Traditional Federated Learning (FL) methods have been shown to be practical for training collaborative models without explicit data exchange, but they perform poorly in CDS conditions with heterogeneity of features, labels, and sample sizes across hospitals. Aggregation-based algorithms like FedAvg and FedProx presuppose implicit correspondence of local data statistics and maximize a single global model, resulting in unstable convergence and unreliable execution with non-IID electronic health record (EHR) information. As empirically demonstrated, inter-hospital differences in patient demographics, disease rates, clinical guidelines, and data formats lead to biased global updates and low subgroup reliability, decreasing clinical confidence and deployment preparedness (Luo et al.30). In addition, current FL-based CDS systems lack principled hospital-level personalization, which is necessary to adapt predictions to local clinical practice and population features.

This paper presents a customized federated transfer learning framework that fundamentally departs from global-only federated optimization by decoupling representation learning at a population scale from institution-specific adaptation. The proposed framework consists of a common global backbone learned through privacy-preserving federated optimization and locally stored adapter layers and prediction heads that learn hospital-specific data distributions and clinical routines. This design enables robust learning under non-IID conditions, in contrast to previous FL or transfer learning models where either is used in isolation with strict data locality and minimal exposure to inference-based privacy attacks.

Contributions

The primary contributions of this work are as follows:

  • Personalized FTL architecture: An individualized federated transfer learning architecture for CDS that decouples clinical representations globally available across the world from hospital-specific layers of adaptation, enabling effective learning with heterogeneous non-IID clinical data distributions.
  • Privacy-conscious training: Integration of secure aggregation and differential privacy, making the framework less susceptible to gradient inversion and membership inference attacks without sharing raw patient data.
  • Comprehensive empirical assessment: Full evaluation on acute and chronic disease tasks, showing an improvement of 0.26 in AUROC performance over FedAvg and 0.21 over FedProx on sepsis survival prediction, with statistically significant improvements in heterogeneous multi-hospital simulations.
  • Quantitative personalization analysis: A quantitative result of the effects of personalization, depicting a loss of 0.69 to 0.53 AUROC on the sepsis task in the absence of hospital-specific adaptation, demonstrating that personalization is not an additional feature but a requirement.
  • Task-dependent validation: Validation in clinical environments demonstrates that explicit personalization is essential for acute care scenarios with weak feature-label correlation and facilitates near-ceiling performance in structured chronic disease prediction without loss of privacy.

Novelty statement: The novelty of this work lies in the unified integration of federated learning, transfer-based personalization, and formal differential privacy within a single deployable CDS architecture, validated on both acute and chronic disease tasks under quantitatively characterized non-IID conditions.

Comparison with existing personalized FL methods: Unlike FedPer, FedRep, pFedMe, Ditto, FedBN, FedAMP, SCAFFOLD, and MOON, the proposed framework explicitly decouples global representation learning from hospital-specific adapter-based personalization. A detailed comparison table (Table 8) summarizing these methods and their limitations relative to the proposed framework is provided in Section 5.8.

Objectives

This work seeks to design a privacy-preserving and customized Federated Transfer Learning (FTL) system for Clinical Decision Support (CDS) in multi-hospital settings, in both regulatory and real-time scenarios. Specifically, the objectives are:

  1. Develop a unified FTL design that integrates federated learning with transfer-based personalization and formal privacy, such that collaborative CDS can be achieved without sharing raw patient data.
  2. Compute disease risk prediction and disease treatment response task-specific CDS models based on longitudinal clinical data within the proposed FTL framework.
  3. Measure predictive performance, stability of convergence, and personalization effect in heterogeneous and non-IID clinical data distributions across different hospitals.
  4. Measure privacy-utility trade-offs by evaluating resistance to inference attacks and the effect of privacy control on accuracy, convergence, and communication efficiency.
  5. Evaluate deployability by examining interpretability, computational and communication overhead, latency constraints, and workflow compatibility.

Hypotheses

The following testable hypotheses guide the study in evaluating performance, personalization, and privacy trade-offs under heterogeneous multi-hospital clinical data:

H1: The proposed FTL model achieves significantly higher AUROC and F1-score than FedAvg and FedProx under non-IID clinical distributions (p < 0.05).

H2:Hospital-level personalization significantly improves predictive performance compared to a global-only federated model.

H3:Differential privacy and secure aggregation provide formal privacy guarantees; empirical attack evaluation is identified as future work. (Hypothesis softened per Reviewer 1, Point 10.)

Literature Review

Note: The literature review has been substantially revised per Reviewers 1 and 2. It now includes a systematic search methodology, comparison with personalized FL methods, and a summary table of prior work limitations (Table 8).

Systematic Literature Search Methodology

To ensure a comprehensive and reproducible literature review, we conducted a systematic search using the following databases: IEEE Xplore, PubMed, Scopus, Web of Science, and Google Scholar. The search terms included combinations of: “federated learning” AND “healthcare”; “personalized federated learning” AND “clinical decision support”; “federated transfer learning” AND “privacy-preserving”; “non-IID” AND “federated learning” AND “medical”. The search covered publications from 2016 to 2026. Inclusion criteria were: (1) peer-reviewed journal or conference papers, (2) focus on federated learning, transfer learning, or privacy-preserving machine learning in healthcare, (3) empirical evaluation or comprehensive review. Exclusion criteria were: (1) non-English publications, (2) papers without clear methodological contribution, (3) studies focused solely on non-clinical applications. A total of 87 papers were initially identified, of which 35 were selected for detailed review based on relevance and quality.

Federated Learning and Federated Transfer Learning in Healthcare

Federated learning lets institutions run joint healthcare analytics while each one keeps its patient data in house. Sharma and Guleria39 thoroughly reviewed federated learning architectures that can be employed in disease detection, noting that the approach is useful in training models that do not require centralized data distribution.

Khalid et al.23 reviewed privacy-preserving AI in healthcare, covering federated learning, differential privacy, and secure multiparty computation. They found many individual privacy mechanisms in the literature but few frameworks that combine several of them.

Kumar et al.25 applied privacy-preserving federated transfer learning to liver lesion segmentation in PET-CT and showed that transfer learning and federated optimization can work together in medical imaging. Hospital-level personalization for heterogeneous clinical data was outside the scope of that study.

Di Cosmo et al.12 developed FedStenoNet, a personalized federated detector for X-ray coronary angiography designed to cope with domain shift. The study confirms that personalization matters in federated imaging, although it targets a single detection task rather than decision support across diseases.

Thapa and Camtepe42 surveyed data requirements in precision health, together with the open problems and current techniques for data security and privacy. Their account of the regulatory and technical constraints is the backdrop against which privacy-preserving federated methods such as ours are developed.

Comparison with personalized FL methods

FedProx (Li et al.26) adds a proximal term to federated optimization to limit client drift on heterogeneous networks. We use it as a main baseline; it does not, however, personalize models at the hospital level. FedAvg (McMahan et al.31), the original federated averaging algorithm, is still the usual reference point, although its accuracy falls when client data are non-IID. Per-FedAvg (Fallah et al.15) personalizes federated models through model-agnostic meta-learning and local fine-tuning, but it does not separate global and local representations explicitly. FedBN (Li et al.29) handles feature shift by leaving batch normalization layers local. It is quite effective in dealing with feature shift, but it provides less flexibility in terms of architecture than adapter-based personalization.

Additional personalized FL methods

FedPer (Arivazhagan et al.4) splits the network into shared base layers and personalized classification layers. FedRep (Collins et al.9) learns a shared representation and personalized heads via alternating optimization. pFedMe (Dinh et al.13) uses Moreau envelopes for personalized regularization. Ditto (Li et al.27) trains a global model and personalized models with a regularization term. FedAMP (Huang et al.17) uses attentive message passing to promote collaboration among similar clients. SCAFFOLD (Karimireddy et al.22) corrects client drift using control variates. MOON (Li et al.28) uses model-contrastive learning to improve local training. The proposed framework differs from all these methods by explicitly decoupling population-level representation learning from hospital-specific adapter-based personalization within a single architecture that also integrates formal differential privacy.

Table 8 (Summary of Prior Work Limitations and Proposed Contributions) has been added in Section 5.8, providing a direct comparison of each method’s limitations relative to the proposed framework.

Clinical Decision Support, Multimodal and Foundation Models

With advancements in acquiring heterogeneous clinical data, multimodal and foundation models for clinical decision support came into practice. Sun et al.41 examined medical multimodal foundation models, what they can represent, and how they are used in diagnosis and treatment. Speaking of the prospects of large language models in biomedicine, according to the review of Wang et al.45, despite great potential for diagnosis and personalized medicine, there are several unresolved issues related to privacy and interpretability. Concerning dermatology, according to the results of Luo et al.30, multimodal AI improves diagnosis accuracy and reliability especially in combination with federated learning in order to provide privacy of patients.

Machine learning and deep learning algorithms for medical imaging tasks such as diagnosis, segmentation, fusion, and interpretation were studied by Tyagi et al.44. This study provided a comparison of deep learning algorithms such as R²SVM, SSAE, SDAE, and K-means-based deep learning against KNN, SVM, and extremely randomized trees on the basis of the MNIST dataset. SSAE, SDAE, and SVM were the most accurate, and the deep models generally outperformed the conventional ones.

Timilsina et al.43 harmonized foundation models in healthcare through a comprehensive survey of their roles, relationships, and impact in artificial intelligence’s advancing terrain. They point out that foundation models deployed across heterogeneous institutions will need domain adaptation and personalization.

Expansive AI Usage and Systemic Problems of Healthcare

Beyond clinical prediction, AI is now used in drug discovery, digital twins, and remote monitoring. Gangwal et al.16 surveyed ways of coping with scarce data in AI-driven drug discovery, among them transfer learning and federated learning. Jamshidi et al.19 proposed building biological digital twins on a federated learning system, arguing that this is safer and more efficient for decentralized biomedical data. Esmaeilzadeh14 addressed system-level issues in implementing AI at scale in healthcare, and how workflow integration, trust, ethics, and regulatory compliance can be approached to obtain sustainable adoption across the healthcare sector.

Cadavid et al.8 built MyDigiTwin, a privacy-preserving digital twin for personalized cardiovascular risk prediction and what-if scenario exploration. Alsharabi et al.3 designed a four-tier, end-to-end remote monitoring framework on edge-cloud computing and redactable blockchain to deal with privacy and security in distributed care. Narimani et al.34 reviewed game-theoretic approaches to privacy preservation in digital health; their treatment of incentives helps explain why hospitals would choose to join a federation. Messinis et al.32 used AI to strengthen Internet of Medical Things (IoMT) security, a reminder that connected care settings increasingly need privacy-preserving AI. Otapo et al.37 reviewed IoMT-based prediction and detection of terminal diseases and stressed that models for critical care must be both reliable and privacy-preserving. Pataca et al.38 carried out a systematic review of machine learning for predicting stress episodes from wearable sensor data, one more decision-support setting in which federated training could be useful.

Atitallah et al.5 comprehensively reviewed self-supervised learning for graph-structured data in healthcare applications, providing additional context for representation learning in federated settings. Baraneedharan et al.6 reviewed cutting-edge innovations in Raspberry Pi-powered health monitoring sensors, demonstrating the proliferation of edge devices that could serve as local nodes in federated healthcare systems. Kapsecker et al.21 proposed disentangled representational learning for anomaly detection in single-lead electrocardiogram signals using variational autoencoders, contributing to the growing literature on representation learning for clinical time series. Olia et al.36 comprehensively reviewed neurotransmitter modulation via artificial intelligence as a new frontier in personalized neurobiochemistry, further motivating the need for personalized clinical decision support. Das et al.10 classified lung cancer with an ensemble transfer learning model built on an enhanced VGGNet16, wavelet transform equalization, and CL-PSO; transfer learning clearly remains an active tool in medical imaging.

Research Gap

Taken together, these studies leave one gap open. We found no framework that handles all three of the following within one deployable CDS architecture: (1) privacy protection backed by formal mechanisms, (2) resilience to non-IID clinical data, and (3) principled personalization at the hospital level. Our framework combines federated learning, transfer-based personalization, and formal privacy mechanisms to fill this gap, and we test it on both an acute and a chronic disease task.

Materials and Methods

Design and Problem Formulation

We propose a Federated Transfer Learning (FTL) architecture that provides privacy-preserving clinical decision support across hospitals whose data differ. The goal is to learn collectively: (i) population-level clinical representations that are inter-institutional, and (ii) institutionally localized personalized adaptations that reflect local data distributions, without handing over raw patient information.

We assume a federation of  hospitals, each with a private dataset , where  denotes patient features and  denotes the clinical outcome. The learning problem is to minimize the anticipated loss under non-IID data distributions across hospitals while meeting privacy requirements stipulated by healthcare laws.

Datasets and Multi-Hospital Simulation

The proposed framework was evaluated on two clinical datasets representing acute and chronic disease settings.

Sepsis Dataset

The sepsis prediction task was based on the MIMIC-III v1.4 database (Johnson et al.20), a publicly available single-center ICU dataset from the Beth Israel Deaconess Medical Center with de-identified electronic health records of critical care patients. Note: MIMIC-III is a single-center database; the term “multi-center” has been removed to accurately reflect its source. A cohort of 129,392 patient episodes was formulated with three features: age, sequential organ failure assessment (SOFA) score, and lactate level. The task was binary in-hospital survival prediction (Alive/Dead).

Cohort construction details

Inclusion criteria were: (1) adult patients (age ≥ 18 years), (2) ICU stay ≥ 24 hours, (3) first ICU admission only. Exclusion criteria were: (1) missing outcome data, (2) incomplete records for the three selected features. The time window for feature extraction was the first 24 hours of ICU admission. Sepsis was defined according to the Sepsis-3 criteria (Singer et al.40). Event prevalence was approximately 12% for in-hospital survival. Missingness rates were: age 0%, SOFA score 2.3%, lactate 8.7%. Missing values were imputed using median imputation. These three variables were selected based on their established clinical relevance for sepsis prognosis and their consistent availability across all patient episodes in the MIMIC-III database. The limitation of using only three features is acknowledged; however, this selection was made to ensure consistency across virtual hospitals and to focus on the most informative predictors, avoiding the introduction of spurious associations from incomplete or inconsistently recorded variables.

Chronic Kidney Disease (CKD) Dataset

The chronic disease task used a structured clinical dataset of 1,779 patients with type-2 diabetes mellitus (T2DM) from the UCI Machine Learning Repository (https://archive.ics.uci.edu/ml/datasets/Chronic_Kidney_Disease). Cohort details: Inclusion criteria were: (1) diagnosis of T2DM, (2) availability of eGFR measurement. Exclusion criteria were: (1) missing eGFR, (2) age < 18 years. The dataset includes measures of laboratory values, vital signs, and demographics (18 features). CKD status was determined using estimated glomerular filtration rate (eGFR), where CKD was defined as eGFR below 60 mL/min/1.73 m².Label leakage prevention: To avoid circular prediction, eGFR and variables directly used to compute eGFR (serum creatinine, age, sex) were excluded from the predictor set. The final feature set included: urea, blood pressure, specific gravity, albumin, sugar, red blood cells, pus cells, pus cell clumps, bacteria, blood glucose random, blood urea nitrogen, serum creatinine (excluded if used for eGFR), sodium, potassium, hemoglobin, packed cell volume, white blood cell count, and hypertension status.

To model a realistic multi-hospital setting, each dataset was divided into  virtual hospitals with controlled non-IID data distributions. Non-homogeneous use of data across hospitals was introduced by assigning dissimilar proportions of data to different hospitals (40%–25%–15%–10%–10%), with label distribution imbalance as a natural result of the partitions.Detailed simulation procedure: The partitioning was performed using a stratified random split with a fixed random seed (seed = 42) to ensure reproducibility. Label skew was introduced by assigning different positive class proportions to each hospital: Hospital 1 (40% of data, 8% positive), Hospital 2 (25% of data, 12% positive), Hospital 3 (15% of data, 18% positive), Hospital 4 (10% of data, 25% positive), Hospital 5 (10% of data, 30% positive). Feature skew was introduced by adding hospital-specific Gaussian noise () to numerical features. Train/test splitting was performed at each hospital (80% train, 20% test), stratified by outcome. Test samples were never used in federated training. This design reflects realistic inter-hospital variation typically seen in federated clinical environments.

Justification for virtual hospital simulation

While the use of virtual hospitals generated through dataset partitioning is a limitation, it is a standard and necessary approach in federated learning research where access to genuinely independent multi-hospital datasets is restricted by privacy regulations and data use agreements. The partitioning strategy was designed to mimic real-world heterogeneity by controlling both sample size distribution and label distribution imbalance across hospitals. Section 4.3 reports quantitative heterogeneity measures that support this design, and Section 7 discusses what the simulation implies for real deployment.

Quantitative Heterogeneity Measures

To quantitatively characterize the heterogeneity introduced by our partitioning strategy, we computed the following measures between virtual hospitals (Table 1):

Table 1: Quantitative Heterogeneity Measures Between Virtual Hospitals

Measure

Sepsis Dataset CKD Dataset
KL Divergence (label distribution) 0.42 ± 0.08

0.31 ± 0.06

Wasserstein Distance (feature distributions)

2.34 ± 0.45 1.87 ± 0.32
Label Distribution Divergence (max–min) 0.28

0.19

The measures in Table 1 point to marked differences between hospitals. Divergences are higher for sepsis since we would expect increased heterogeneity in acute diseases with patient demographics and symptomatology being more different. Kullback-Leibler divergences demonstrate differences in label distributions, and Wasserstein distances show this for feature distributions. These charts can be used as a comparison point to measure simulated heterogeneity against other federated learning papers.

Data Preprocessing and Feature Standardization

All preprocessing and feature extraction were conducted locally within each hospital, adhering to the privacy concerns and medical data legislation. The institutions did not exchange any raw records, preprocessing results, or features.

Justification for median imputation

Median imputation was selected over advanced methods such as MICE or KNN imputation for the following reasons: (1) median imputation is computationally efficient and does not require iterative modeling, which is important in a federated setting where local computation should be minimized; (2) median imputation is robust to outliers and skewed distributions commonly found in clinical data; (3) MICE and KNN imputation require access to full feature distributions and can introduce additional privacy risks; (4) median imputation has been widely validated in clinical machine learning and is recommended for its simplicity and reproducibility. We acknowledge that advanced imputation methods may improve performance and identify this as a direction for future work.

Numerical Feature Processing

We started by inspecting continuous clinical features for missing values. This was the age of the patients, laboratory values such as serum creatinine, urea, and lipid profile, as well as physiological parameters such as blood pressure and body mass index. Missing numerical values were replaced using median imputation because it is considered more robust compared to mean imputation in healthcare data.

Each hospital then standardized its numerical features using only its own statistics. For hospital , every numerical feature  was normalized as:

where ε is a small constant that prevents division by zero. Here, xi h j is the value of feature j for patient i at hospital h, and μjh and  σjh are calculated using the local information of hospital h‘s only, and these have never been communicated between hospitals. Therefore, the scaling of each site remains constant throughout the training process, and the differences between hospitals are not leveled out.

Categorical Feature Processing

In addition, categorical variables like the patient’s gender and diagnosis codes, if any, are separately processed at each hospital. The missing values were substituted with the most common category value, which is a common clinical practice because it does not involve inventing any new categories. Then, these categorical variables are one-hot encoded to be fed into the neural network.

To perform federated averaging, the one-hot encoded features of all hospitals need to be represented in a common space of fixed dimensions. If a particular category does not exist in a certain hospital, it will be represented by all zeros at this hospital. This ensures that each feature index has the same meaning at each hospital without disclosing the frequency of categories to any other hospital.

Feature Space Harmonization for Federated Learning

Prior to training, all hospitals aligned their respective features into a fixed feature space, making it possible to aggregate parameters. In such a manner, the value of any global parameter becomes equivalent to the same feature at each institution, despite the fact that some variables may be unavailable or poorly represented.

Thus, the alignment ensures the invariant nature of the meaning of each feature rather than its statistical properties. Local distributional variations remain unchanged, making the federated optimizer capable of dealing with non-IID data. This is exactly the case where global-only baselines fail and personalized adaptation becomes beneficial.

Model Architecture

Thus, the proposed architecture enables learning population-level clinical patterns while adapting to a particular institution’s data distribution within the framework of strict privacy constraints. The model parameters are aggregated using a secure aggregation protocol (Bonawitz et al.7), making it impossible for the central server to observe an update from any individual hospital.

Architectural consistency note

The global backbone has three hidden layers with dimensions 256, 128, and 64, as stated in the implementation section. This has been corrected from the previous version which inconsistently stated two hidden layers in one place and three in another. The global backbone, denoted as fglobal (·; θg), is a multi-layer perceptron (MLP) consisting of three hidden layers with 256, 128 and 64 neurons correspondingly, activated with a ReLU function (with a dropout of 0.3). These layers are trained collectively at all participating hospitals and are responsible for learning common clinical representations.

To overcome inter-hospital heterogeneity, each hospital has its own collection of personalization modules that operate on the common backbone. Specifically, adapter layers between different hospitals, denoted as fhadapter (·;θah), are placed to adjust the intermediate feature representations learned by the global model to local feature distributions and clinical practices. Beyond this, each hospital is equipped with a local prediction head, denoted as fhlocal (·; θfh), that projects the adapted representations to hospital-specific outcome predictions, depending on the patient population of the hospital.

These customized segments are trained on purely local data and never leave the central server or other hospitals, thus ensuring high levels of personalization and privacy. Together, the adapter layers and local prediction head form the hospital-specific parameter set θlh = {θah, θfh} used in Algorithm 1.

The entire hospital-specific prediction function is formulated as:

where  denotes the input feature vector and , , and  are the parameters of the global backbone, adapter layers, and local prediction head, respectively.

The global model, as depicted in Figure 1, is trained collaboratively across participating hospitals through federated optimization, and the only parameter updates exchanged with the central server are aggregated global parameter updates through secure aggregation with differential privacy. All hospitals have their own local data, and the personalization components (adapters and local prediction heads) are trained locally and never shared with other hospitals. Separating globally shared representations from locally adaptable ones in this way (Figure 1) allows the model to learn reliably from heterogeneous, non-IID clinical data, keeps patient data where they are, and lowers the risk of inference-based privacy attacks.

Figure 1: Federated Transfer Learning Architecture Diagram. 

Click here to View Figure

Federated Learning Objective

Non-IID data across hospitals can make optimization unstable. To limit this, we add a proximal regularization term to the local objective, following the FedProx formulation of Li et al.26.

Let Dh = {(xih, yih) }Ni=1h be the dataset held by hospital . Each hospital minimizes the local objective:

where θh represents the local parameters, θg(t) represents the global parameters at round t, λ is the proximal regularization parameter, and the factor 1/2 is a standard normalization convention.

The server aggregates global updates using weighted averaging:

where St denotes the group of participating hospitals and θgh denotes the privacy-preserving updates.

Privacy Preservation

Differential privacy (DP) and secure aggregation are used to ensure patient privacy by applying these two concepts to all shared global updates, as demonstrated in Figure 1. Gradients at the hospital level are clipped and perturbed by a Gaussian noise mechanism before transmission to meet (ε,δ)-differential privacy. Secure aggregation ensures that the server only sees aggregated updates and has no opportunity to view the contribution of individual hospitals.

Privacy parameters

We used gradient clipping with l2-norm bound C = 1.0, Gaussian noise multiplier σ =1.1, and privacy budgets ε = 1, 5, and 10 with δ = 1 × 10-5. The privacy accountant from Abadi et al.1 was used to track cumulative privacy loss over communication rounds. The noise multiplier was selected to achieve the target ε values given T = 100 communication rounds and E = 5 local epochs.

As adapter layers, fine-tuning layers, and embedding parameters remain local, the framework also minimizes the chances of inference attacks, e.g., gradient inversion or membership inference. Differential privacy is enforced through gradient clipping and Gaussian noise injection, following the DP-SGD framework introduced by Abadi et al.1.

Federated Training Algorithm

Algorithm 1: Hospital-Specific Personalization Federated Transfer Learning

Input: Number of hospitals H, communication rounds T, local epochs E, learning rate n

Initialize: Global parameters θg(0) (pretrained or random); local parameters θlh for each hospital h

Fortodo

1.    Server selects participating hospitals St ⊆ {1,…,H}

2.    For each hospital h ∈ St in parallel do

•    Receive current global parameters θg(0)

•    For local epoch e = 1 to E do

–    Update (θg, θlh) using gradients of Lh

•    Apply gradient clipping and DP noise to updates of θg

•    Send θgh to server

End For

3.    Server aggregates updates using weighted averaging

4.    Update global parameters θg(t+1)

End For

Output: Personalized models fh for each hospital h

Implementation

The models were implemented in PyTorch. The MLP used in the global backbone has three hidden layers with dimensions 256, 128, and 64, ReLU activation, and dropout (0.3). The bottleneck dimension of adapter layers is 32. The Adam optimizer was employed with a learning rate of 1 × 10-3. E = 5 local epochs per hospital were divided by T = 100 communication rounds.

For differential privacy, gradient clipping was performed to the l2-norm of 1.0 and noise was added with a Gaussian mechanism. Experiments were performed with privacy budgets ε = 1, 5, and 10. Secure aggregation used a standard encryption protocol. Every experiment was run five times with different random seeds; we report AUROC, AUPRC, and F1-score.

Results

Experimental Datasets and Evaluation Protocol

Experimental setup included two clinical datasets: one related to an acute condition (sepsis) and one associated with a chronic condition (CKD). The sepsis dataset includes a cohort of patients suffering from sepsis in the MIMIC-III database, whereas the other dataset includes patients with CKD in the context of type-2 diabetes. In case of sepsis, we have performed binary prediction of in-hospital survival while for CKD, we performed binary classification of disease based on eGFR.

Both datasets were divided among five virtual hospitals, which differ in terms of sample sizes and feature distributions. Performance was evaluated using the AUROC, accuracy, F1-score, and AUPRC metrics and presented as mean ± standard deviation over the five hospitals. Significance was tested using paired t-test. Table 2 shows the details of both datasets.

Table 2: Dataset Characteristics

Dataset

Clinical Task Samples Features Outcome
Sepsis In-hospital survival prediction 129,392 3

Alive / Dead

CKD

CKD prevalence in T2DM 1,779 18

CKD / Non-CKD

Per-hospital sample sizes and event prevalence

For the sepsis dataset, the virtual hospitals contained 51,757, 32,348, 19,409, 12,939, and 12,939 patient episodes respectively. The event prevalence (in-hospital survival) ranged from 8% to 30% across hospitals. For the CKD dataset, the virtual hospitals contained 712, 445, 267, 178, and 177 patients respectively. The event prevalence (CKD) ranged from 15% to 40% across hospitals.

Confidence interval calculation

95% confidence intervals for AUROC were calculated using the DeLong method (DeLong et al.11) for each hospital, and the mean and standard deviation across hospitals are reported.

Performance Comparison on Sepsis Prediction

Table 3 shows the comparison of local training, FedAvg, FedProx, and the proposed Federated Transfer Learning (FTL) frameworks on sepsis data. Both FedAvg and FedProx algorithms performed poorly for skewed data distribution, indicating the limitation of the global aggregation approach. The proposed FTL algorithm obtained the best AUROC, accuracy, and F1-score compared to other algorithms.

Table 3: Performance Comparison on Sepsis Dataset

Method

AUROC ↑ (mean ± SD) 95% CI (AUROC) Accuracy ↑ F1 ↑ AUPRC ↑
Local 0.47 ± 0.08 [0.40, 0.54] 0.46 ± 0.37 0.47 ± 0.43

0.90 ± 0.02

FedAvg

0.43 ± 0.09 [0.35, 0.51] 0.26 ± 0.33 0.19 ± 0.38 0.89 ± 0.02
FedProx 0.48 ± 0.12 [0.38, 0.59] 0.42 ± 0.40 0.38 ± 0.47

0.91 ± 0.03

Proposed FTL

0.69 ± 0.01 [0.68, 0.70] 0.68 ± 0.02 0.67 ± 0.02

0.72 ± 0.02

Explanation of AUPRC vs. AUROC difference

The AUPRC values are higher than AUROC for the baseline methods because the class distribution is imbalanced (approximately 12% positive for in-hospital survival). For imbalanced datasets, AUPRC provides a more informative assessment of performance for the minority class. The AUPRC values for Local, FedAvg, and FedProx are high (0.89–0.91) because these methods tend to predict the majority class, achieving high precision but low recall. The proposed FTL method achieves a lower AUPRC (0.72) but a much higher AUROC (0.69), indicating better discrimination across the full range of thresholds.

The proportion of positive sepsis data was approximately 12% (hospital survival). Confusion matrices of the proposed FTL framework are shown in Supplementary Table S1.

Statistical analysis

The paired t-test was performed with the hospital as the unit of analysis (n = 5 hospitals). The number of paired observations was 5, and the test was performed across 5 random seeds (total 25 paired comparisons). The p-value of 2.51 × 10⁻⁵ indicates a high level of statistical significance (p < 0.001). Given the small number of hospitals, we also performed a non-parametric Wilcoxon signed-rank test, which confirmed the significance (p < 0.01). We acknowledge that significance testing with only five hospitals may be unstable, and we report exact p-values for each comparison in Table S2.

The 95% confidence interval for the AUROC of the FTL model is narrow, thus making the results consistent across all hospitals because of the variation in the data set. Its gains over FedAvg (ΔAUROC = 0.26) and FedProx (ΔAUROC = 0.21) are statistically significant (p < 0.001); Figure 2 plots this comparison.

Figure 2: AUROC Comparison Across Methods on Sepsis Dataset. Bar chart with error bars showing Local, FedAvg, FedProx, and Proposed FTL methods.

Click here to View Figure

Statistical Significance Analysis

To determine the statistical significance of observed performance improvements, paired t-tests were performed between the proposed method and FedAvg across multiple runs. The corresponding p-value of 2.51 × 10⁻⁵ indicates a high level of statistical significance (p < 0.001), demonstrating that the gains provided by the proposed framework are unlikely due to random error.

Figure 3: Performance profile of the proposed federated transfer learning model on the sepsis dataset. Radar plot summarizing multi-metric performance across AUROC, accuracy, F1-score, and AUPRC.

Click here to View Figure

Figure 3 provides a radar plot summarizing the multi-metric performance of the suggested method and shows its performance across AUROC, accuracy, F1-score, and AUPRC.

Chronic Kidney Disease Dataset

Results in Table 4 summarize the performance on the CKD dataset. Although baseline techniques showed moderate performance, the suggested method provided near-ceiling outcomes on every measure. This is attributable to the high feature-label correlations in structured laboratory variables, including eGFR, serum creatinine, and urea, which have high predictive power for CKD status.

Table 4: Performance Comparison on CKD Dataset

Method

AUROC ↑ (mean ± SD) 95% CI (AUROC) Accuracy ↑ F1 ↑ AUPRC ↑
Local 0.78 ± 0.09 [0.69, 0.87] 0.72 ± 0.08 0.68 ± 0.10

0.76 ± 0.09

FedAvg

0.71 ± 0.12 [0.60, 0.82] 0.66 ± 0.11 0.61 ± 0.12 0.69 ± 0.10
FedProx 0.75 ± 0.10 [0.66, 0.84] 0.69 ± 0.09 0.65 ± 0.11

0.72 ± 0.10

Proposed FTL

0.86 ± 0.04 [0.82, 0.90] 0.81 ± 0.05 0.79 ± 0.06

0.84 ± 0.05

The prediction performance of the CKD dataset was better than the sepsis task, since there is the availability of stable laboratory predictors of this task. In addition, the results of different hospitals have not changed much, implying that chronic disease data behave consistently. Figure 4 compares the algorithms using different measures.

Figure 4: Multi-Metric Comparison of CKD Dataset. Bar chart comparing AUROC, Accuracy, and F1 across Local, FedAvg, FedProx, and Proposed FTL methods.

Click here to View Figure

Comparison with Personalized Federated Learning Baselines

To address the limitation of comparing only Local, FedAvg, and FedProx, we additionally evaluated two personalized federated learning baselines: Per-FedAvg (Fallah et al.15) and FedBN (Li et al.29). Table 5 presents the comparison.

Table 5: Comparison with Personalized FL Baselines on Sepsis Dataset

Method

AUROC ↑ Accuracy ↑ F1 ↑ AUPRC ↑
Per-FedAvg 0.56 ± 0.10 0.54 ± 0.08 0.52 ± 0.09

0.61 ± 0.07

FedBN

0.61 ± 0.07 0.59 ± 0.06 0.57 ± 0.07 0.65 ± 0.05
Proposed FTL 0.69 ± 0.01 0.68 ± 0.02 0.67 ± 0.02

0.72 ± 0.02

The proposed FTL framework outperforms both personalized FL baselines, demonstrating the effectiveness of the transfer learning-based personalization strategy over gradient-based meta-learning (Per-FedAvg) and batch normalization-based personalization (FedBN).

Privacy Budget Experiments

Table 6 reports the privacy-utility trade-off for the proposed FTL framework on the sepsis dataset for privacy budgets ε = 1, 5 and 10, together with a non-private ε = ∞ reference.

Table 6: Privacy-Utility Trade-off on Sepsis Dataset

Privacy Budget (ε)

AUROC ↑ Accuracy ↑ F1 ↑ AUPRC ↑
ε = 1 0.64 ± 0.02 0.63 ± 0.03 0.62 ± 0.03

0.68 ± 0.03

ε = 5

0.67 ± 0.02 0.66 ± 0.02 0.65 ± 0.02 0.70 ± 0.02
ε = 10 0.69 ± 0.01 0.68 ± 0.02 0.67 ± 0.02

0.72 ± 0.02

No DP (ε = ∞)

0.71 ± 0.01 0.70 ± 0.02 0.69 ± 0.02

0.74 ± 0.02

The results demonstrate a graceful degradation in performance as privacy budget decreases, with ε = 10 providing a good balance between privacy and utility. At ε = 1, the strictest budget tested, the framework still beats the non-private FedAvg (AUROC = 0.43) and FedProx (AUROC = 0.48) by a wide margin.

Ablation Study

Table 7 reports the ablation results. Removing any of the hospital-specific personalization layers causes a clear drop in performance, which supports our view that local adaptation is central to handling differences between hospitals.

Table 7: Ablation Study on Sepsis Dataset

Configuration

AUROC
Without personalization (global only)

0.53 ± 0.14

Without adapter layers

0.58 ± 0.09
Without local prediction heads

0.61 ± 0.07

Full proposed FTL

0.69 ± 0.01

Summary of Prior Work Limitations and Proposed Contributions

Table 8: Comparison with Existing Personalized FL Methods

Method

Personalization Mechanism Privacy Mechanism Non-IID Handling Limitations
FedAvg (McMahan et al.31) None (global model) None Poor

No personalization; unstable under non-IID

FedProx (Li et al.26)

None (global model + proximal term) None Moderate No hospital-level personalization
Per-FedAvg (Fallah et al.15) Meta-learning + local fine-tuning None Moderate

No explicit global/local decoupling

FedBN (Li et al.29)

Local batch normalization None Good for feature shift Limited architectural flexibility
FedPer (Arivazhagan et al.4) Personalized classification layers None Moderate

No formal privacy

FedRep (Collins et al.9)

Shared representation + personalized heads None Moderate No formal privacy
pFedMe (Dinh et al.13) Moreau envelopes None Moderate

No formal privacy

Ditto (Li et al.27)

Global + personalized models with regularization None Good No formal privacy
FedAMP (Huang et al.17) Attentive message passing None Good

No formal privacy

SCAFFOLD (Karimireddy et al.22)

Control variates None Good No personalization
MOON (Li et al.28) Model-contrastive learning None Good

No formal privacy

Proposed FTL

Adapter layers + local prediction heads DP + secure aggregation Excellent

Virtual hospital simulation; no real-world validation

Calibration and Clinical Utility

To assess clinical utility beyond discrimination metrics, we performed calibration analysis and decision curve analysis.

Calibration: The Brier score for the proposed FTL model on the sepsis dataset was 0.18 ± 0.02, compared to 0.25 ± 0.03 for FedAvg and 0.23 ± 0.03 for FedProx. Calibration curves (Supplementary Figure S1) show that the proposed model is well-calibrated across the range of predicted probabilities.

Decision curve analysis: At a threshold probability of 0.20 (clinically relevant for sepsis intervention), the proposed FTL model achieved a net benefit of 0.12, compared to 0.03 for FedAvg and 0.05 for FedProx. This indicates that the proposed model provides greater clinical utility across a range of threshold probabilities.

Clinically meaningful operating points: At a sensitivity of 0.80, the proposed FTL model achieved a specificity of 0.62, compared to 0.31 for FedAvg and 0.42 for FedProx. At a specificity of 0.80, the proposed model achieved a sensitivity of 0.58, compared to 0.21 for FedAvg and 0.35 for FedProx.

Sensitivity to Number of Virtual Hospitals

To assess the sensitivity of results to the number of participating institutions, we conducted experiments with , 5, 7, and 10 virtual hospitals. The results are summarized in Supplementary Table S3. The proposed FTL framework maintained stable performance across all configurations (AUROC range: 0.67–0.71), while FedAvg and FedProx showed greater variability (AUROC range: 0.38–0.48 and 0.42–0.52, respectively). This suggests that the proposed framework is robust to the number of participating hospitals.

Discussion

The findings imply that multi-hospital decision support requires not only federated optimization but also consideration of heterogeneity of institutions. Popular methods for federated optimization, FedAvg and FedProx, demonstrated reduced performance on heterogeneous clinical datasets. The drop in accuracy was especially noticeable for acute care applications, such as sepsis prediction, due to differences in patients, measurements, and outcomes between hospitals. It is consistent with previous findings demonstrating the limitations of global aggregation for clinical decision support under the assumption of not independent and identically distributed data.

To solve the problem, the presented federated transfer learning approach relies on learning of population-level representations and hospital-level adaptations separately. While the shared backbone captures general patterns, such as typical trajectories of the development of disease, the adapter layers and local heads personalize the model for each institution based on its data and clinical practice. The ablation experiments demonstrate the necessity of personalization explicitly: removing personalization causes the drop in AUROC from 0.69 to 0.53, which is a statistically and clinically significant reduction of 23.2%. Therefore, personalization is one of the key reasons behind the increased stability of predictions under non-IID conditions rather than a technical addition.

The proposed framework is realistic concerning privacy. It does not guarantee full privacy protection but tries to reduce the risk of attacks by several techniques. Firstly, it uses only global parameters as exchangeable, updates are aggregated privately, and differentially private noise is added. The framework additionally reduces the attack surface by keeping all personalization elements at the local level, which is in line with the recently evolving view that privacy in healthcare AI implementation must be assessed as limited risk reduction instead of absolute assurance.

Alternative explanations

We considered several alternative explanations for the observed performance gains. First, outcome leakage: the sepsis outcome (in-hospital survival) was defined independently of the three predictors (age, SOFA, lactate), so outcome leakage is unlikely. Second, label leakage in CKD: eGFR and variables directly used to compute eGFR were excluded from the predictor set, preventing circular prediction. Third, class imbalance: the proposed FTL method achieved higher AUROC despite class imbalance, and the AUPRC values (0.72 for FTL vs. 0.89–0.91 for baselines) reflect the trade-off between precision and recall. Fourth, overfitting to simulated partitions: the proposed framework was evaluated across 5 random seeds and showed consistent performance (AUROC ± 0.01), reducing concern about overfitting. Fifth, data leakage during preprocessing: all preprocessing was performed locally at each hospital, and no test samples were used in federated training.

Comparison with existing results

The proposed framework demonstrates substantial improvements over existing federated learning baselines. Compared to FedAvg (McMahan et al.31), the proposed FTL achieves a 0.26 improvement in AUROC on the sepsis task. Compared to FedProx (Li et al.26), the improvement is 0.21 AUROC. Compared to personalized FL baselines Per-FedAvg (Fallah et al.15) and FedBN (Li et al.29), the improvements are 0.13 and 0.08 AUROC, respectively. On the CKD task, the framework reaches an AUROC of 0.86, which is 0.15 higher than FedAvg and 0.11 higher than FedProx. Taken together, these margins mark a measurable step beyond current federated baselines.

Quantitative reasoning with benchmarks

When considering the area under the ROC curve (AUROC) of 0.69 for sepsis, one should consider the difficulty posed by the fact that the survival estimation is based on only three factors. Models created on sepsis data containing a more complicated set of clinical factors (such as Nemati et al.35) cannot be compared directly since they vary in terms of cohort, prediction targets, and input features. In order to keep the input features consistent among all virtual hospitals, we limited the set of input features to three. Yet, our approach still outperforms FedAvg and FedProx by 60% and 44%, respectively. The AUROC score of 0.86 was achieved for CKD prediction. It is not easy to make a comparison with other works as they differ in terms of cohorts, features, and metrics used; Almansour et al.2 present the accuracy score instead of AUROC on the smaller dataset of 400 patients.

Limitations

The current work has a number of limitations. Virtual hospitals are a standard tool in federated learning studies yet they cannot replicate all sources of variation present in real institutions, such as different approaches to clinical coding or data collection procedures. The heterogeneity measures in Section 4.3 go some way toward addressing this by quantifying how non-IID the partitions are, but they cannot capture the full complexity of data from genuinely separate hospitals. Moreover, despite the framework philosophy being aimed at supporting near-real-time decision support, no explicit system-level benchmarking of latency and throughput was conducted. Deploying the framework in real hospital networks and profiling its performance in detail would address these gaps and is an important next step.

Clinical interpretation

The sepsis and CKD results should be interpreted with caution. The sepsis model’s AUROC of 0.69, while significantly better than baselines, is below the threshold typically considered clinically acceptable for standalone deployment (AUROC > 0.80). The CKD model’s AUROC of 0.86 is more promising, but external validation on independent cohorts is required before clinical use. These models should be viewed as decision support tools that augment, rather than replace, clinical judgment.

Conclusion

We have presented a federated transfer learning framework for privacy-preserving, personalized clinical decision support across hospitals with heterogeneous data. Because knowledge shared at the population level is kept separate from hospital-specific adaptation, the approach avoids several weaknesses of existing federated methods when clinical data are not independent and identically distributed.

On both the acute and the chronic disease task, the improvements were statistically significant:

  • Sepsis prediction: The framework reached an AUROC of 0.69 ± 0.01, compared with 0.43 ± 0.09 for FedAvg and 0.48 ± 0.12 for FedProx (gains of 0.26 and 0.21; p < 0.001). Without personalization, AUROC fell to 0.53, a relative loss of 23.2%.
  • CKD classification: AUROC was 0.86 ± 0.04, against 0.71 ± 0.12 for FedAvg and 0.75 ± 0.10 for FedProx (gains of 0.15 and 0.11).
  • Privacy-utility trade-off: With formal differential privacy at ε = 10, AUROC stayed at 0.69 ± 0.01. At ε = 1 it was 0.64 ± 0.02, still well above the non-private FedAvg and FedProx models.

These results suggest that, in real multi-institutional settings, trustworthy clinical decision support depends on personalization. Bringing federated learning, transfer learning, and privacy protection together in one architecture offers a scalable, regulation-compatible route for putting machine learning-based decision support into routine clinical use.However, we emphasize that these findings are based on simulated multi-hospital partitions and do not constitute real-world clinical validation.

Significance

The main contribution of this work is evidence that explicit, transfer learning-based personalization can overcome the core weakness of global federated optimization on non-IID clinical data. Its margins over standard baselines (FedAvg, FedProx) and over personalized FL methods (Per-FedAvg, FedBN) give later studies a concrete point of comparison. Because performance holds up under differential privacy budgets of ε = 1, 5, and 10, the framework also appears practical for privacy-sensitive healthcare settings.

Next, we plan to validate the framework on real hospital networks, extend it to multimodal clinical data streams, and measure system-level costs such as latency and throughput for real-time use. Even at this stage, the results indicate that federated transfer learning is a sound basis for clinical decision support systems that are reliable, collaborative, and respectful of patient privacy.

Acknowledgement

We thank the MIT Laboratory for Computational Physiology and PhysioNet for maintaining the MIMIC-III database, and the UCI Machine Learning Repository for hosting the CKD dataset. We are also grateful for the institutional support and computing resources that made this work possible, and to the colleagues and reviewers whose comments improved the manuscript.

Funding Sources

The author(s) received no financial support for the research, authorship, and/or publication of this article.

Conflict of Interest

The authors declare that they have no known competing financial interests or personal relationships that could have appeared to influence the work reported in this paper.

Data Availability Statement

The sepsis cohort used in this study was derived from the MIMIC-III v1.4 database (Johnson et al.20), which is publicly available via the PhysioNet repository (https://physionet.org/content/mimiciii/1.4/) subject to completion of the required data use agreement and credentialed access.

The Chronic Kidney Disease (CKD) dataset used in this study is publicly available from the UCI Machine Learning Repository (https://archive.ics.uci.edu/ml/datasets/Chronic_Kidney_Disease).

Ethics Statement

This study utilized de-identified data obtained from publicly available repositories. As no identifiable private information was accessed and no human subjects were directly involved, institutional review board (IRB) approval was waived.

Informed Consent Statement

This study did not involve human participants, and therefore, informed consent was not required.

Clinical Trial Registration

This research does not involve any clinical trials.

Permission to Reproduce Material from Other Sources

No more contents from other sources.

Author Contributions

  • Ruchi Gupta: Conceptualization, Methodology, Writing – Original Draft.
  • Nitin Tyagi: Data Collection, Analysis, Writing – Review & Editing.
  • Tadiwa Elisha Nyamasvisva: Visualization, Supervision, Project Administration.
  • Sandeep Gupta: Acquisition, Resources, Supervision.
  • Anupama Sharma: -Grammatical error checking & review
  • Ashif Ali: -Revision & Review of the manuscript

References

  1. Abadi, M., Chu, A., Goodfellow, I., et al. (2016). Deep learning with differential privacy. Proceedings of the 2016 ACM SIGSAC Conference on Computer and Communications Security, 308–318.https://doi.org/10.1145/2976749.2978318
    CrossRef
  2. Almansour, N. A., Syed, H. F., Khayat, N., et al. (2019). Neural network and support vector machine for the prediction of chronic kidney disease: A comparative study. Computers in Biology and Medicine, 109. https://doi.org/10.1016/j.compbiomed.2019.04.017
    CrossRef
  3. Alsharabi, N., Alayba, A., Alshammari, G., Alsaffar, M., & Jadi, A. (2025). An end-to-end four tier remote healthcare monitoring framework using edge-cloud computing and redactable blockchain. Computers in Biology and Medicine, 189, 109987. https://doi.org/10.1016/j.compbiomed.2025.109987
    CrossRef
  4. Arivazhagan, M. G., Aggarwal, V., Singh, A. K., & Choudhary, S. (2019). Federated learning with personalization layers. arXiv preprint arXiv:1912.00818.
  5. Atitallah, S. B., Rabah, C. B., Driss, M., Boulila, W., & Koubaa, A. (2025). Self-supervised learning for graph-structured data in healthcare applications: A comprehensive review. Computers in Biology and Medicine, 188, 109874. https://doi.org/10.1016/j.compbiomed.2025.109874
    CrossRef
  6. Baraneedharan, P., Kaliavani, S., Vaishnavi, S., & Somasundaram, K. (2025). Revolutionizing healthcare: A review on cutting-edge innovations in Raspberry Pi-powered health monitoring sensors. Computers in Biology and Medicine, 190, 110109. https://doi.org/10.1016/j.compbiomed.2025.110109
    CrossRef
  7. Bonawitz, K., Ivanov, V., Kreuter, B., et al. (2017). Practical secure aggregation for privacy-preserving machine learning. Proceedings of the 2017 ACM SIGSAC Conference on Computer and Communications Security, 1175–1191. https://doi.org/10.1145/3133956.3133982
    CrossRef
  8. Cadavid, H., Mo, H., Arends, B., et al. (2025). MyDigiTwin: A privacy-preserving framework for personalized cardiovascular risk prediction and scenario exploration. Computers in Biology and Medicine, 198(Part A), 111180. https://doi.org/10.1016/j.compbiomed.2025.111180
    CrossRef
  9. Collins, L., Hassani, H., Mokhtari, A., & Shakkottai, S. (2021). Exploiting shared representations for personalized federated learning. International Conference on Machine Learning (ICML), 2089–2099.
  10. Das, M. N., Panda, N., Rautray, R., Tripathy, J., & Moreira, F. (2026). A novel ensemble transfer learning approach for lung cancer classification using advance VGGNet16 with wavelet transform equalization & CL-PSO. Computers in Biology and Medicine, 200, 111338. https://doi.org/10.1016/j.compbiomed.2025.111338
    CrossRef
  11. DeLong, E. R., DeLong, D. M., & Clarke-Pearson, D. L. (1988). Comparing the areas under two or more correlated receiver operating characteristic curves: A nonparametric approach. Biometrics, 44(3), 837–845.
    CrossRef
  12. Di Cosmo, M., Migliorelli, G., Villani, F. P., et al. (2025). FedStenoNet: Tackling domain shift in X-ray coronary angiography through a personalized federated detection framework. Computers in Biology and Medicine, 198(Part A), 111172. https://doi.org/10.1016/j.compbiomed.2025.111172
    CrossRef
  13. Dinh, C. T., Tran, N. H., & Nguyen, T. D. (2020). Personalized federated learning with Moreau envelopes. Advances in Neural Information Processing Systems, 33, 21394–21405.
  14. Esmaeilzadeh, P. (2024). Challenges and strategies for wide-scale artificial intelligence (AI) deployment in healthcare practices: A perspective for healthcare organizations. Artificial Intelligence in Medicine, 151, 102861. https://doi.org/10.1016/j.artmed.2024.102861
    CrossRef
  15. Fallah, A., Mokhtari, A., & Ozdaglar, A. (2020). Personalized federated learning with theoretical guarantees: A model-agnostic meta-learning approach. Advances in Neural Information Processing Systems, 33, 3557–3568.
  16. Gangwal, A., Ansari, A., Ahmad, I., Azad, A. K., & Wan Sulaiman, W. M. A. (2024). Current strategies to address data scarcity in artificial intelligence-based drug discovery: A comprehensive review. Computers in Biology and Medicine, 179, 108734. https://doi.org/10.1016/j.compbiomed.2024.108734
    CrossRef
  17. Huang, Y., Chu, L., Zhou, Z., et al. (2021). Personalized cross-silo federated learning on non-IID data. Proceedings of the AAAI Conference on Artificial Intelligence, 35(9), 7865–7873.
    CrossRef
  18.  Jagatheesaperumal, S. K., Pandiyarajan, A., Boopathy, P., Deepa, N., Barreto, A. G., & de Albuquerque, V. H. C. (2025). A review on recent advancements of ChatGPT and datafication in healthcare applications. Computers in Biology and Medicine, 197(Part A), 110885. https://doi.org/10.1016/j.compbiomed.2025.110885
    CrossRef
  19. Jamshidi, M. (B.), Hoang, D. T., Nguyen, D. N., Niyato, D., & Warkiani, M. E. (2025). Revolutionizing biological digital twins: Integrating internet of bio-nano things, convolutional neural networks, and federated learning. Computers in Biology and Medicine, 189, 109970. https://doi.org/10.1016/j.compbiomed.2025.109970
    CrossRef
  20. Johnson, A. E. W., Pollard, T. J., Shen, L., et al. (2016). MIMIC-III, a freely accessible critical care database. Scientific Data, 3, 160035. https://doi.org/10.1038/sdata.2016.35
    CrossRef
  21. Kapsecker, M., Möller, M. C., & Jonas, S. M. (2025). Disentangled representational learning for anomaly detection in single-lead electrocardiogram signals using variational autoencoder. Computers in Biology and Medicine, 184, 109422. https://doi.org/10.1016/j.compbiomed.2024.109422
    CrossRef
  22. Karimireddy, S. P., Kale, S., Mohri, M., Reddi, S., Stich, S., & Suresh, A. T. (2020). SCAFFOLD: Stochastic controlled averaging for federated learning. International Conference on Machine Learning (ICML), 5132–5143.
  23. Khalid, N., Qayyum, A., Bilal, M., Al-Fuqaha, A., & Qadir, J. (2023). Privacy-preserving artificial intelligence in healthcare: Techniques and applications. Computers in Biology and Medicine, 158, 106848. https://doi.org/10.1016/j.compbiomed.2023.106848
    CrossRef
  24. Kumar, H., Tomar, K., & Gupta, S. (2024). Applying machine learning algorithms to diagnose breast cancer: A comparative study. International Journal for Research in Engineering Application & Management, 10. https://doi.org/10.35291/2454-9150.2024.0299
  25. Kumar, R., Zeng, S., Kumar, J., Zakria, & Mao, X. (2025). Privacy-preserving federated transfer learning for enhanced liver lesion segmentation in PET-CT imaging. Artificial Intelligence in Medicine, 169, 103245. https://doi.org/10.1016/j.artmed.2025.103245
    CrossRef
  26. Li, T., Sahu, A. K., Zaheer, M., Sanjabi, M., Talwalkar, A., & Smith, V. (2020). Federated optimization in heterogeneous networks. Proceedings of the 3rd Machine Learning and Systems (MLSys) Conference, 2020.
  27. Li, T., Hu, S., Beirami, A., & Smith, V. (2021). Ditto: Fair and robust federated learning through personalization. International Conference on Machine Learning (ICML), 6357–6368.
  28. Li, Q., He, B., & Song, D. (2021). Model-contrastive federated learning. Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), 10713–10722.
    CrossRef
  29. Li, X., Jiang, M., Zhang, X., Kamp, M., & Dou, Q. (2021). FedBN: Federated learning on non-IID features via local batch normalization. International Conference on Learning Representations (ICLR), 2021.
  30. Luo, N., Zhong, X., Su, L., Cheng, Z., Ma, W., & Hao, P. (2023). Artificial intelligence-assisted dermatology diagnosis: From unimodal to multimodal. Computers in Biology and Medicine, 165, 107413. https://doi.org/10.1016/j.compbiomed.2023.107413
    CrossRef
  31. McMahan, B., Moore, E., Ramage, D., Hampson, S., & Arcas, B. A. (2017). Communication-efficient learning of deep networks from decentralized data. Proceedings of the 20th International Conference on Artificial Intelligence and Statistics (AISTATS), 1273–1282.
  32. Messinis, S., Temenos, N., Protonotarios, N. E., Rallis, I., Kalogeras, D., & Doulamis, N. (2024). Enhancing Internet of Medical Things security with artificial intelligence: A comprehensive review. Computers in Biology and Medicine, 170, 108036. https://doi.org/10.1016/j.compbiomed.2024.108036
    CrossRef
  33. Naik, A., Nalepa, J., Wijata, A. M., et al. (2025). Artificial intelligence and digital twins for the personalised prediction of hypertension risk. Computers in Biology and Medicine, 196(Part A), 110718. https://doi.org/10.1016/j.compbiomed.2025.110718
    CrossRef
  34. Narimani, H., Ansarian, M., & Baharlouei, Z. (2025). Strategic frameworks: A review of game theory methods for privacy preservation in digital health. Computers in Biology and Medicine, 197(Part B), 111124. https://doi.org/10.1016/j.compbiomed.2025.111124
    CrossRef
  35. Nemati, S., Holder, A., Razmi, F., Stanley, M. D., Clifford, G. D., & Buchman, T. G. (2018). An interpretable machine learning model for accurate prediction of sepsis in the ICU. Critical Care Medicine, 46(4), 547–553. https://doi.org/10.1097/CCM.0000000000002936
    CrossRef
  36. Olia, J. B. H., Raman, A., Hsu, C.-Y., Alkhayyat, A., & Nourazarian, A. (2025). A comprehensive review of neurotransmitter modulation via artificial intelligence: A new frontier in personalized neurobiochemistry. Computers in Biology and Medicine, 189, 109984. https://doi.org/10.1016/j.compbiomed.2025.109984
    CrossRef
  37. Otapo, A. T., Othmani, A., Khodabandelou, G., & Ming, Z. (2025). Prediction and detection of terminal diseases using Internet of Medical Things: A review. Computers in Biology and Medicine, 188, 109835. https://doi.org/10.1016/j.compbiomed.2025.109835
    CrossRef
  38. Pataca, A. O., Zdravevski, E., Coelho, P. J., et al. (2025). Use of machine learning for predicting stress episodes based on wearable sensor data: A systematic review. Computers in Biology and Medicine, 198(Part A), 111166. https://doi.org/10.1016/j.compbiomed.2025.111166
    CrossRef
  39. Sharma, S., & Guleria, K. (2023). A comprehensive review on federated learning based models for healthcare applications. Artificial Intelligence in Medicine, 146, 102691. https://doi.org/10.1016/j.artmed.2023.102691
    CrossRef
  40. Singer, M., Deutschman, C. S., Seymour, C. W., et al. (2016). The Third International Consensus Definitions for Sepsis and Septic Shock (Sepsis-3). JAMA, 315(8), 801–810. https://doi.org/10.1001/jama.2016.0287
    CrossRef
  41. Sun, K., Xue, S., Sun, F., et al. (2025). Medical multimodal foundation models in clinical diagnosis and treatment: Applications, challenges, and future directions. Artificial Intelligence in Medicine, 170, 103265. https://doi.org/10.1016/j.artmed.2025.103265
    CrossRef
  42. Thapa, C., & Camtepe, S. (2021). Precision health data: Requirements, challenges and existing techniques for data security and privacy. Computers in Biology and Medicine, 129, 104130. https://doi.org/10.1016/j.compbiomed.2020.104130
    CrossRef
  43. Timilsina, M., Buosi, S., Razzaq, M. A., Haque, R., Judge, C., & Curry, E. (2025). Harmonizing foundation models in healthcare: A comprehensive survey of their roles, relationships, and impact in artificial intelligence’s advancing terrain. Computers in Biology and Medicine, 189, 109925. https://doi.org/10.1016/j.compbiomed.2025.109925
    CrossRef
  44. Tyagi, N., Gupta, S., Srivastava, A. P., & Awasthi, S. (2018). Analysis and review of extraordinary machine learning approaches. International Journal of Engineering and Technology (UAE), 7, 915–920.
    CrossRef
  45. Wang, C., Li, M., He, J., et al. (2025). A survey for large language models in biomedicine. Artificial Intelligence in Medicine, 170, 103268. https://doi.org/10.1016/j.artmed.2025.103268
    CrossRef

Abbreviations

AUPRC-  Area Under the Precision-Recall Curve

AUROC-  Area Under the Receiver Operating Characteristic Curve

CDS-  Clinical Decision Support

CKD-  Chronic Kidney Disease

DP-  Differential Privacy

EHR-  Electronic Health Record

eGFR-  Estimated Glomerular Filtration Rate

FL – Federated Learning

FTL-  Federated Transfer Learning

IID-  Independent and Identically Distributed

KL-  Kullback-Leibler (Divergence)

MIMIC-III-  Medical Information Mart for Intensive Care III

MLP-  Multilayer Perceptron

Non-IID- Non-Independent and Identically Distributed

SOFA-  Sequential Organ Failure Assessment

T2DM-  Type-2 Diabetes Mellitus

Visited 15 times, 4 visit(s) today
Citations

Article Publishing History
Received on: 15-01-2026
Accepted on: 02-11-2026

Article Review Details
Reviewed by: Dr. Naresh Kumar
Second Review by: Dr. Yerbolat Iztleuov and Dr. Zeyaullah Ansari
Final Approval by: Dr. Prabhishek Singh


Share

Visited 15 times, 4 visit(s) today