Hybrid Federated and Split Learning for Privacy Preserving Clinical Prediction and Treatment Optimization

Farzana Akter; Rakib Hossain; Deb Kanna Roy Toushi; Mahmood Menon Khan; Sultana Amin; Lisan Al Amin

cs.LGFeb 17, 2026

Hybrid Federated and Split Learning for Privacy Preserving Clinical Prediction and Treatment Optimization

Authors: Farzana Akter, Rakib Hossain, Deb Kanna Roy Toushi, Mahmood Menon Khan, Sultana Amin, Lisan Al Amin

Organizations: Dept. of Information Technology, Washington University of Science and Technology, Alexandria, VA, USA · Cognitive Links LLC, Maryland, USA

Abstract

Collaborative clinical decision support is often constrained by governance and privacy rules that prevent pooling patient-level records across institutions. We present a hybrid privacy-preserving framework that combines Federated Learning (FL) and Split Learning (SL) to support decision-oriented healthcare modeling without raw-data sharing. The approach keeps feature-extraction trunks on clients while hosting prediction heads on a coordinating server, enabling shared representation learning and exposing an explicit collaboration boundary where privacy controls can be applied. Rather than assuming distributed training is inherently private, we audit leakage empirically using membership inference on cut-layer representations and study lightweight defenses based on activation clipping and additive Gaussian noise. We evaluate across three public clinical datasets under non-IID client partitions using a unified pipeline and assess performance jointly along four deployment-relevant axes: factual predictive utility, uplift-based ranking under capacity constraints, audited privacy leakage, and communication overhead. Results show that hybrid FL-SL variants achieve competitive predictive performance and decision-facing prioritization behavior relative to standalone FL or SL, while providing a tunable privacy-utility trade-off that can reduce audited leakage without requiring raw-data sharing. Overall, the work positions hybrid FL-SL as a practical design space for privacy-preserving healthcare decision support where utility, leakage risk, and deployment cost must be balanced explicitly.

Figures & tables

Paradigm	Exchanged Objects	Post-2020 Examples	Decision-Support Gap
Federated Learning (FL)	Model updates; optional secure aggregation	Challenges [ 2 ] ; healthcare evidence [ 3 ] ; EHR review [ 8 ]	Often predictive-only; leakage auditing is inconsistent in applied pipelines [ 7 ] .
Split Learning (SL)	Cut-layer activations and backpropagated gradients	SL survey [ 15 ] ; representation attacks [ 13 , 14 ]	Privacy/compute focus; decision utility such as uplift prioritization is rarely evaluated.
Hybrid FL and SL	Split representations with federated aggregation	SplitFed [ 4 ] ; edge analysis [ 5 ] ; PPSFL [ 10 ]	Limited treatment-optimization focus; unified auditing and cross-dataset validation are often incomplete.
Causal / Uplift (Centralized)	Centralized $(X,T,Y)$ with counterfactual estimators	CATE challenges [ 6 ] ; HTA review [ 16 ] ; uplift benchmarking [ 17 ]	Typically assumes pooled data; governance and multi-site constraints are not addressed.
This Work	Hybrid FL and SL with uplift evaluation	Proposed framework	Unified utility, communication, and membership inference auditing across datasets and client partitions.

TABLE I: Taxonomy of recent collaborative learning in healthcare and the gap addressed in this work.

Dataset	$T$ (treatment proxy)	$Y$ (outcome)	Client partition ( client_id )
eICU-demo [ 20 ]	Early vasopressor exposure in first 6 hours (binary)	In-hospital mortality (binary)	Admission diagnosis groups (top diagnoses vs other), inducing site-like heterogeneity
MEPS 2022 [ 21 ]	Statin exposure from prescribed medicine records (binary)	Inpatient utilization indicator (binary)	Region-based partition to mimic policy and access differences
NHANES 2015–2018 [ 22 ]	Statin mention in prescription file (binary)	Controlled cholesterol indicator (binary)	Survey cycle-based partition (cycle I vs J) for temporal shift

TABLE II: Unified uplift construction across datasets.

T

denotes a treatment or exposure proxy;

Y

is the outcome; client_id defines a non-IID collaborative split.

Mode	Exchange	Computation placement	Key rationale
Centralized	None (pooled data)	Single trainer	Upper-bound utility reference
FL (FedAvg)	Model weights/updates	Full model trained locally	Keeps raw data local; aggregates updates [ 2 ]
SL	Activations and gradients at cut layer	Trunk on client, head on server	Reduces client compute; shifts depth to server
Hybrid FL-SL	Split activations plus federated aggregation of client trunks	Trunk local (federated), head server (split)	Combines parallelism and partitioning [ 4 , 5 ]

TABLE III: Compared training modes. “Exchange” indicates the primary object communicated across clients and server.

Fig. 1 : End-to-end research framework for privacy-preserving collaborative uplift modeling.

Client trunk	2-layer MLP: $d\rightarrow 64\rightarrow 32$ (ReLU)
Cut layer	After trunk output $z\in\mathbb{R}^{32}$ (Split/Hybrid)
Server heads	Two independent heads: $32\rightarrow 1$ + sigmoid
Training	Adam; $\eta_{c}=\eta_{s}=10^{-3}$ ; batch $=256$ (train), $512$ (eval); local $E=1$
Defense	Activation clipping $c=1.0$ ; Gaussian noise $\sigma=0.05$

TABLE IV : Reproducibility settings used in all experiments. Training rounds

R

follow Table V .

Dataset	Method	AUROC	AUUC	End	Trim	Worst	Comm	Rounds	MIA
				Uplift	%	AUUC	(MB)		AUC
eicu	Centralized	$0.51$	$-0.03$	$-0.29$	$9.92$	$-0.81$	$0.00$	$5$	N/A
eicu	FedAvg	$0.48$	$-0.11$	$-0.29$	$9.92$	$-0.37$	$2.04$	$8$	N/A
eicu	Split	$0.51$	$-0.13$	$-0.29$	$9.92$	$-0.72$	$3.61$	$4$	0.49
eicu	Hybrid	$0.49$	$-0.10$	$-0.29$	$9.92$	$-0.06$	$3.81$	$4$	0.48
eicu	Hybrid+Pers.	$0.54$	$-0.29$	$-0.29$	$9.92$	$-0.71$	$3.81$	$4$	0.50

TABLE V : Main results across datasets. AUROC: factual discrimination. AUUC: uplift ranking under trimming (may be negative with observational proxies). Worst-client AUUC: non-IID robustness. Comm and rounds: deployment cost; MIA AUC: leakage audit for split/hybrid modes.

Panel of Figure 7 — Fig. 2 : Uplift curves under capacity targeting (top fraction by predicted uplift).

Fig. 3 : Privacy–utility trade-off: AUUC vs membership inference AUC for defended configurations.

Configuration		AUUC across clients			AUROC across clients
Dataset	Method	Mean	Std	Worst	Mean	Std	Worst
eicu	Centralized	$-0.4650$	$0.3288$	$-0.7938$	$0.3730$	$0.1674$	$0.2056$
eicu	FedAvg	$-0.2521$	$0.2334$	$-0.4855$	$0.3045$	$0.1045$	$0.2000$
eicu	Split	$-0.3475$	$0.0180$	$-0.3655$	$0.3167$	$0.2056$	$0.1111$
eicu	Hybrid	$-0.4282$	$0.0871$	$-0.5153$	$0.4844$	$0.1177$	$0.3667$
eicu	Hybrid+Pers.	$-0.2799$	$0.1645$	$-0.4444$	$0.3276$	$0.2443$	$0.0833$

TABLE VI : Non-IID robustness across clients. Mean and standard deviation are computed across clients; worst-client metrics highlight conservative deployment risk.

Uplift Modeling

See all 12 papers

A Privacy-Preserving Federated Learning Method for Heterogeneous Clients
Nov 8, 2025Yusaku Negoya, Feifei Cui, Zilong Zhang +3Privacy-Preserving Federated Learning Differentially Private Federated Learning
Department of Computer and Network Engineering, The University of Electro-Communications, Tokyo, Japan · School of Computer Science and Technology, Hainan University, Haikou 570228, China. · Department of Electrical and Computer Engineering, University of Houston, Houston, TX, 77204, USA +1
Privacy-Preserving Federated Learning via Differential Privacy and Homomorphic Encryption for Cardiovascular Disease Risk Modeling
Apr 30, 2026Gaurang Sharma, Juha Pajula, Aada Illikainen +4Homomorphic Encryption Privacy-Preserving ML
VTT Technical Research Centre of Finland Ltd, 33720, Tampere, Finland · VTT Technical Research Centre of Finland Ltd, 02150, Espoo, Finland · Solita Oy, 00100 Helsinki, Finland +1
HybridFL: A Federated Learning Approach for Financial Crime Detection
Feb 22, 2026Afsana Khan, Marijn ten Thij, Guangzhi Tang +1Anti-Money Laundering Federated Learning
Department of Advanced Computing Sciences, Maastricht University, Netherlands · Department of Cognitive Science and Artificial Intelligence, Tilburg University, Netherlands

Nov 8, 2025J/K move · Enter open · S save

A Privacy-Preserving Federated Learning Method for Heterogeneous Clients

Yusaku Negoya, Feifei Cui, Zilong Zhang, Miao Pan+2

Department of Computer and Network Engineering, The University of Electro-Communications, Tokyo, Japan · School of Computer Science and Technology, Hainan University, Haikou 570228, China. · Department of Electrical and Computer Engineering, University of Houston, Houston, TX, 77204, USA+1

Privacy-Preserving Federated Learning Differentially Private Federated Learning Heterogeneous Federated Learning

Open paper AskPDF arXiv

Federated Learning (FL) with Differential Privacy (DP) can ensure the protection of healthcare data privacy against malicious user attacks. However, FL with the DP method faces an inherent trade-off: stronger privacy protection degrades predictive accuracy due to injected noise. On the other hand, Homomorphic Encryption (HE) enables computations on encrypted data and the aggregation of encrypted gradients without introducing DP-induced noise, thereby potentially increasing the predictive accuracy. However, it may increase the computation cost. To improve the predictive accuracy while considering the computational ability of heterogeneous clients, we propose a Privacy-Preserving Machine Learning (PPML)-Hybrid method. In the proposed PPML-Hybrid method, clients distributedly select either HE or DP based on their computational resources, so that HE clients contribute noise-free updates while DP clients reduce computational overhead. Meanwhile, clients with high computational resources can flexibly adopt HE or DP according to their privacy needs. Performance evaluation on omics datasets shows that our proposed method achieves comparable predictive accuracy while significantly reducing computation time relative to HE-only. Additionally, it outperforms DP-only methods under equivalent or stricter privacy budgets.

Abstract

Uplift Modeling

Explore similar work

A Privacy-Preserving Federated Learning Method for Heterogeneous Clients

Privacy-Preserving Federated Learning via Differential Privacy and Homomorphic Encryption for Cardiovascular Disease Risk Modeling

HybridFL: A Federated Learning Approach for Financial Crime Detection