A newly developed decentralized machine learning framework named FL-HFS has demonstrated that multi-institutional clinical systems can collaboratively predict cardiovascular disease with high diagnostic precision while keeping patient medical records strictly confidential. Cardiovascular conditions remain the primary cause of mortality across the globe, yet training predictive algorithms across diverse hospital networks has historically encountered severe legal, ethical, and computational obstacles. The newly introduced architecture resolves these constraints by unifying local three-stage feature selection with an adaptive federated optimization protocol across distributed data sources. Evaluated across 5,430 patient records gathered from six distinct cardiovascular cohorts, the system attained an overall classification accuracy of 88.3 percent alongside a balanced F1-score of 91 percent. Rather than transmitting raw health data to a centralized repository, each participating healthcare site conducts localized training and exchanges solely model parameters. This configuration provides a scalable and privacy-preserving paradigm for clinical risk detection, matching or closely approaching the diagnostic effectiveness of conventional centralized predictive models without compromising institutional data sovereignty.
The global urgency driving this research stems from the massive public health burden imposed by cardiovascular disorders, which claim between 17 million and 18 million lives every year and account for approximately one-third of all global deaths. Conditions encompassing coronary artery disease, heart failure, arrhythmias, and stroke severely impair cardiac function and require early, highly accurate detection to reduce fatality rates through timely clinical interventions. While classical risk factors including hypertension, diabetes, smoking, physical inactivity, obesity, and chronic stress accelerate disease progression, standard centralized computing models face prohibitive barriers in aggregating medical information. Stringent data privacy regulations, specifically the General Data Protection Regulation and the Health Insurance Portability and Accountability Act, restrict the cross-border and inter-institutional transfer of sensitive electronic health records. Furthermore, transmitting massive datasets from Internet of Medical Things devices and clinical facilities introduces prohibitive bandwidth costs, vulnerability to unauthorized data breaches, and systemic scaling bottlenecks, emphasizing the necessity of distributed computational methodologies in modern medicine.
To navigate the limitations of centralized health analytics, the investigators structured their methodology around federated learning, an advanced paradigm where participating institutional clients train models collaboratively while retaining all local patient observations on internal servers. Under this workflow, a central coordination server initializes global model parameters and broadcasts them across all designated hospital nodes. Each local participant conducts private model training using only its native records, adjusting network weights without exposing diagnostic reports, physiological waveforms, or patient histories. Following localized optimization, the nodes transmit exclusively their mathematical updates, such as gradient vectors or network hyperparameters, back to the central server. The global server subsequently merges these discrete client updates into an improved global architecture using specialized proximal algorithms before redistributing the refined network for subsequent communication cycles. This repetitive aggregation loop continues until optimization converges, providing continuous knowledge sharing across regionally dispersed facilities while providing rigorous structural confidentiality against unauthorized patient data extraction.
A critical technical innovation introduced within this architecture is an independent, three-stage hybrid feature selection pipeline executed locally at each client node prior to model training. Clinical datasets frequently contain redundant, uninformative, or noisy measurements that elevate computational complexity and degrade generalizability across diverse demographic groups. To isolate the most informative diagnostic indicators, each hospital site sequentially applies variance thresholding, mutual information ranking, and L1-regularized sparse regression. The initial variance filter removes features exhibiting a variance below a threshold of 0.01 after localized feature scaling, successfully eliminating static or non-informative attributes. Next, mutual information metrics measure the non-linear dependency between each retained attribute and the binary disease outcome to rank feature significance. Finally, an L1-penalized logistic regression model shrinks inconsequential feature coefficients to zero, retaining a sparse subset of clinically significant variables. Executing this feature selection routine strictly on local training partitions ensures model interpretability, minimizes dimensional overhead, and prevents cross-client statistical leakage.
The research team assembled an experimental framework encompassing six distinct clinical heart disease datasets to simulate real-world healthcare environments, totaling 5,430 patient records. These data sources included the Framingham Heart Study along with the Cleveland, Hungary, Switzerland, VA Long Beach, and Statlog cohorts obtained from established clinical repositories. Thirteen primary clinical variables formed the baseline feature space, comprising age, sex, chest pain category, resting blood pressure, serum cholesterol, electrocardiographic results, peak heart rate, exercise-induced angina, and ST depression measurements. Each node applied a uniform localized preprocessing pipeline, imputing missing values via training-partition medians, clipping extreme measurement outliers within the 0.5th to 99.5th percentile boundary, and executing Z-score normalization. Differences in column naming conventions across sources, such as Framingham variable definitions, were systematically aligned with standard clinical schemas. To resolve internal class distribution skews without causing cross-client contamination, the authors applied synthetic minority over-sampling methods exclusively to local training partitions, preserving the natural distributions of validation and testing sets.
A profound hurdle in decentralized healthcare computation is the prevalence of non-independent and identically distributed data across institutions, which the study categorized into three distinct operational shifts. First, client nodes exhibited pronounced label distribution skew, with positive cardiovascular disease frequency varying dramatically from 15.2 percent in the Framingham cohort to 93.5 percent in the Switzerland dataset. Second, systematic variations in feature missingness existed; while certain cohorts fully recorded fluoroscopy and thallium scintigraphy indicators, others omitted them entirely. Third, the study highlighted semantic label heterogeneity, where five cohorts defined positive cases through angiographic arterial narrowing exceeding 50 percent, whereas the Framingham cohort recorded ten-year coronary heart disease events. To overcome this divergence, the investigators incorporated a proximal stability parameter within the FedProx framework to prevent local models from straying too far from global updates. Furthermore, they formulated an area-under-the-curve-weighted aggregation mechanism that weights client contributions based on discriminative reliability, preventing nodes with pathological class balances from degrading global parameter convergence.
The underlying machine learning architecture consisted of a deep neural network featuring multiple fully connected layers, residual connections, batch normalization, Gaussian Error Linear Unit activation functions, and dropout regularization. This deep structure was specifically configured to capture complex, non-linear relationships embedded across multi-institutional diagnostic measurements. Experiments were conducted utilizing an NVIDIA GeForce RTX 3060 graphics processor equipped with 12.9 gigabytes of memory, accelerating computation through Automatic Mixed Precision and unified parallelized execution. Local optimization employed the AdamW algorithm paired with a cosine annealing learning rate scheduler and binary cross-entropy and focal loss functions to stabilize convergence under imbalanced conditions. The training routine operated over 80 global communication rounds, with each participating client completing 20 local training epochs per communication iteration using an empirically verified batch size of 32 and a base learning rate of 0.001 determined via systematic grid searches.
Comparative experimental assessments demonstrated that the proposed federated framework achieved an overall accuracy of 88.3 percent, demonstrating superior computational efficiency and communication economy compared to alternative decentralized models. For example, traditional federated averaging and TabNet architectures reached approximately 82.2 percent accuracy, whereas complex optimization models based on artificial bee colony algorithms and matched averaging attained accuracies between 92.8 percent and 93.8 percent. However, those alternative models required between 2,408 and 3,018 communication rounds to achieve convergence, creating massive network traffic, severe latency, and elevated energy consumption across distributed medical endpoints. In stark contrast, the newly designed architecture reached stable convergence above 88 percent in only 80 communication rounds. While centralized machine learning models like K-nearest neighbors and ensemble stacking reached accuracies of 93 percent and 91 percent respectively, their mandatory reliance on centralized raw data aggregation makes them impractical under modern medical privacy constraints.
Detailed diagnostic evaluation confirmed the balanced clinical utility of the system, which generated an overall precision score of 0.90, a recall rate of 0.92, and an area under the receiver operating characteristic curve between 0.91 and 0.93. An analysis of the global confusion matrix across a representative experimental fold revealed 160 true positive diagnoses and 574 true negative classifications, alongside 72 false negatives and 256 false positive identifications. The investigators noted that this diagnostic distribution reflects an intentional, sensitivity-prioritized optimization strategy designed for clinical screening environments. In cardiovascular triage protocols, missing an actual positive disease manifestation carries catastrophic clinical consequences, whereas false positive classifications can be systematically clarified through secondary non-invasive diagnostics. Local client accuracy trajectories also demonstrated progressive cross-institutional learning, with performance across individual centers rising steadily over consecutive communication cycles to reach localized accuracies ranging between 82 percent and 98 percent across different hospital partitions.
Rigorous statistical validation through paired two-tailed t-tests across ten independent experimental runs initialized with unique random seeds confirmed that the performance advantages of the framework were statistically significant with p-values below 0.05. Furthermore, an extensive ablation investigation revealed that omitting the hybrid feature selection module caused a sharp drop of approximately six percentage points in diagnostic accuracy, proving that filtering noisy and redundant biological variables is essential for preventing client-level overfitting. The researchers concluded that the framework successfully balances predictive efficacy, institutional privacy, and computational overhead under extreme clinical data heterogeneity. Future research will explore expanding the federated architecture to large-scale cross-silo healthcare networks, integrating dynamic client participation protocols to accommodate intermittent network connectivity, and embedding explainable artificial intelligence techniques to provide medical practitioners with transparent, interpretable visual rationales for automated risk predictions.
Subject of Research: Decentralized federated learning and hybrid clinical feature selection for privacy-preserving cardiovascular disease prediction
Article Title: Federated learning framework with hybrid feature selection for heart disease prediction
Article References: Manita, Mahajan, G., Gupta, N. K., & Chauhan, G. S. (2026). Federated learning framework with hybrid feature selection for heart disease prediction. Discover Computing, 29(1), Article 863. https://doi.org/10.1007/s10791-026-10658-0
Image Credits: AI Generated
DOI: 10.1007/s10791-026-10658-0
Keywords: Federated Learning, Heart Disease Prediction, Cardiovascular Disease, Machine Learning, Privacy-Preserving AI, Feature Selection, Medical Informatics, Digital Health, Federated, learning, framework, hybrid
News Source: Frances Kline. (November 9, 2026). Federated Learning Framework Enhances Privacy and Accuracy in Heart Disease Prediction. Scienmag.



