Federated Data Imputation

The core constraint is that raw patient data with missing values cannot leave the local site. Imputation models must be trained collaboratively using only aggregate statistics, model parameters, or synthetic data.

• Federated EM Algorithm: Sites compute local sufficient statistics, which are aggregated to update global parameter estimates for likelihood-based imputation.
• Secure Aggregation: Homomorphic encryption or secure multi-party computation masks individual contributions during the imputation model's training phase.
• Differential Privacy: Noise is injected into shared statistics to prevent membership inference attacks that could reveal which patients had missing values.

Missing data mechanisms—Missing Completely at Random (MCAR), Missing at Random (MAR), and Missing Not at Random (MNAR)—can vary drastically across hospitals due to different clinical workflows, equipment, and patient demographics.

• A tertiary care center may have MNAR data for a specific biomarker because it's only measured in critically ill patients.
• A community clinic might have MAR data where missingness in a lab test is predicted by the patient's age, which is fully observed.
• Federated imputation models must learn site-specific missingness masks or use federated domain adaptation to avoid biased global estimates.

Many federated imputation methods extend Multiple Imputation by Chained Equations (MICE) to a decentralized setting. Each variable with missing values is imputed in a round-robin fashion using all other variables as predictors.

• Local Modeling: Each site trains a regression model for the target variable on its complete cases.
• Parameter Aggregation: Model coefficients are securely averaged across sites.
• Stochastic Imputation: Sites draw from the predictive posterior distribution to create multiple plausible imputed datasets, preserving uncertainty.
• This iterative process continues until convergence across the federated network.

Generative models like Variational Autoencoders (VAEs) and Generative Adversarial Networks (GANs) can be trained in a federated manner to learn the joint distribution of clinical data, including missingness patterns.

• A federated VAE learns a latent representation of patient data across sites. Missing values are imputed by sampling from the conditional distribution given observed features.
• Federated GAIN (Generative Adversarial Imputation Nets) uses a generator to impute missing values and a discriminator to distinguish real from imputed data, with both trained via federated averaging.
• These methods naturally handle complex, non-linear relationships and can generate multiple imputations to reflect uncertainty.

Clinical sites often have different feature spaces—one hospital may collect a genetic panel that another does not. Federated imputation must handle this vertical partition of data.

• Federated Vertical Imputation: Uses overlapping patient cohorts (linked via encrypted identifiers) to learn cross-feature relationships without sharing raw data.
• Transfer Imputation: A model trained on a feature-rich source client is adapted to impute missing features in a target client with a narrower feature set.
• Federated Matrix Factorization: Decomposes the global patient-feature matrix into latent factors, with each site holding a partition of the factor matrices.

Imputation can introduce or amplify bias. Federated systems must provide transparent audit trails and fairness guarantees without accessing raw data.

• Federated Imputation Diagnostics: Sites compute local diagnostics (e.g., imputation variance, coverage rates) that are aggregated to assess global imputation quality.
• Fairness Constraints: Optimization objectives can incorporate group fairness metrics to ensure imputation accuracy is consistent across demographic subgroups.
• Provenance Tracking: Blockchain or verifiable logs record which imputation model version and parameters were used for each patient record, critical for regulatory compliance.

Understanding why data is missing is critical before imputation. The pattern of missingness in federated clinical data falls into three categories:

• Missing Completely at Random (MCAR): Absence is unrelated to any variable. A blood sample was lost in transit.
• Missing at Random (MAR): Absence is related to observed data. Older patients are less likely to complete a lifestyle questionnaire.
• Missing Not at Random (MNAR): Absence depends on the unobserved value itself. Patients with severe symptoms skip follow-up appointments.

In federated settings, the missingness mechanism may itself be non-IID across sites, requiring site-specific imputation strategies.

A distributed adaptation of the classic MICE framework that generates multiple plausible imputed datasets to reflect uncertainty:

• Each variable with missing data is imputed using a conditional model (e.g., regression, random forest) trained on all other variables.
• In federated MICE, these conditional models are trained collaboratively across sites using federated averaging.
• Multiple imputation cycles produce m complete datasets; analysis is performed on each and results are pooled using Rubin's rules.

This method properly accounts for imputation uncertainty, which single imputation methods ignore, and is robust to label distribution skew across clinical sites.

A latent factor approach that decomposes a partially observed data matrix into low-rank representations, filling missing entries through collaborative filtering principles:

• The global data matrix X is approximated as the product of two lower-dimensional matrices: U (client embeddings) and V (feature embeddings).
• Each client locally updates its row of U using its observed data, while V is aggregated and shared across the network.
• Missing values are predicted as the dot product of the corresponding client and feature vectors.

This technique is particularly effective for high-dimensional clinical data with structured missingness patterns, such as incomplete lab panels across hospital networks.

Imputing missing values inherently risks leaking information about the original missing data. Differential privacy (DP) provides formal guarantees:

• Epsilon (ε): A privacy budget parameter controlling the trade-off between accuracy and privacy. Lower ε means stronger privacy.
• Gaussian Mechanism: Adding calibrated noise to imputation model updates before aggregation masks individual contributions.
• Privacy Accounting: Tracking cumulative privacy loss across multiple imputation rounds using Rényi DP or moments accountant.

A key challenge is that imputation quality degrades under strong DP guarantees, requiring careful privacy-utility trade-off analysis for clinical applications.

Using deep generative architectures—such as Variational Autoencoders (VAEs) and Generative Adversarial Networks (GANs)—to learn the joint data distribution across sites and sample plausible values for missing entries:

• Federated VAEs: Each client trains a local encoder/decoder; latent distributions are aggregated to learn a global generative model without sharing raw data.
• Federated GAIN: A GAN-based architecture where a generator imputes missing values and a discriminator attempts to distinguish real from imputed entries, trained collaboratively.
• These methods capture complex non-linear relationships and multimodal distributions that simpler regression-based imputation misses.

Particularly valuable for imputing missing medical imaging metadata or genomic markers with intricate correlation structures.