MedFlexFusion: Adaptive Medical Data Fusion

Updated 3 July 2026

MedFlexFusion is a framework for flexibly fusing heterogeneous medical modalities using context-sensitive, adaptive mechanisms.
It employs techniques like PerM-MoE and two-stage attentional fusion to dynamically weight inputs and handle missing data effectively.
The approach yields robust performance in clinical applications, including neuroimaging, cardiac analysis, and EHR time series tasks.

MedFlexFusion refers to a class of architectures and algorithms specifically designed to enable flexible, robust fusion of heterogeneous medical data modalities—be they imaging, tabular clinical measurements, signals, or time series—under challenging regimes of missingness or varying input availability. This paradigm encompasses a spectrum of technically distinct but philosophically aligned frameworks such as PerM-MoE for neuroimaging (Burns et al., 8 Sep 2025), the MedFlexFusion module in cardiac analysis (Zhang et al., 18 Aug 2025), multiplicative fusion for irregular time series (Hsieh et al., 12 Nov 2025), hierarchical fusion in FlexCare (Xu et al., 2024), and diffusion-based models for arbitrary-input image fusion (Xu et al., 11 Sep 2025). All share the core goal of learning flexible, context-sensitive representations that adaptively combine available modalities for downstream clinical tasks.

1. Problem Motivation and Conceptual Overview

Traditional multimodal medical models fall short when faced with high rates of missing modalities during inference—a defining characteristic in real-world clinical data, such as incomplete neuroimaging, periodic labs, or asynchronous time series. Classic fusion mechanisms (early or late fusion, additive embedding, or simple concatenation) are either rigid to missingness or discard cross-modal interactions. MedFlexFusion frameworks address these deficiencies by equipping models with learned mechanisms to handle:

Arbitrary subsets of available modalities, without retraining or model surgery.
Cross-modal, value- and context-sensitive weighting, supporting dynamic task-adaptive integration.
High-dimensional, heterogeneous representations (e.g., 3D imaging, tabular clinical, free text) for a wide range of predictive or generative analyses.

This approach is motivated both by clinical reality—where missing data is endemic and the prognostic value of any given modality is highly context-, patient-, and task-dependent—and by empirical findings that naïve early or late fusion yields substantial performance degradation under high missingness (Burns et al., 8 Sep 2025, Zhang et al., 18 Aug 2025, Hsieh et al., 12 Nov 2025, Xu et al., 2024, Xu et al., 11 Sep 2025).

2. Technical Methodologies and Architectures

MedFlexFusion methods span multiple technical instantiations, each tailored to the nature of the target data modalities and clinical tasks.

Sparse Mixture-of-Experts with Per-Modality Routers (PerM-MoE)

PerM-MoE (Burns et al., 8 Sep 2025) replaces a single, global router with independent gating networks per modality. Each 3D imaging modality (T1 MRI, FLAIR MRI, amyloid PET, tau PET) is encoded via a modality-specific 3D-CNN. The embeddings—observed or imputed from a missing-modality bank—are routed through per-modality MoE blocks:

Each router $g_i$ computes gating logits $v^i$ over $E$ experts ( $E=16$ ; select top- $k=2$ per modality).
Self-attention is applied across modalities before routing.
Outputs are concatenated and fed to a final MLP prediction head alongside a baseline covariate (e.g., prior CDR-SB). This decoupled routing provides sharper expert specialization and increased robustness to missing modalities.

Two-Stage Attentional Fusion with Adaptive Gating

The MedFlexFusion module (Zhang et al., 18 Aug 2025) implements a two-stage strategy:

Local cross-modal attention fuses each modality's encoding with a shared, input-adaptive key/value bank, producing context-aware locally fused vectors.
Global gating aggregates all local fusions alongside unimodal features via element-wise sigmoid gates, generating the final multimodal embedding. All mixing coefficients (λ-weights in the bank, gates $G_i$ ) are learned end-to-end and updated per-instance, enabling real-time adaptation to task and data availability.

Multiplicative Embedding Fusion for Clinical Time Series

In irregular, asynchronous settings (as with EHR time series), MedFuse introduces the multiplicative fusion operator (Hsieh et al., 12 Nov 2025): $\mathbf{e}_{f,v} = \mathbf{e}_f \odot \mathbf{e}_v$ where $\mathbf{e}_f$ is the feature-identity embedding and $\mathbf{e}_v$ is a value-conditioned embedding. This design preserves feature identity while allowing the observed value to nonlinearly modulate the feature representation. Each observed measurement is tokenized as $(f,v,t)$ and fed sequentially to a Transformer encoder. Missing modality and measurement pairs are simply omitted, requiring no imputation.

Task- and Modality-Aware Hierarchical Fusion

FlexCare (Xu et al., 2024) employs hierarchical attention and an MoE-refinement module:

Unimodal encoders and a Transformer-based intra/inter-modality encoder yield both individual modality and cross-modal combination embeddings.
A task-guided router, conditioned jointly on task and modality context, dynamically selects expert subnetworks and aggregates their outputs via hierarchical attention. This supports both flexible input configurations and multitask output heads.

Diffusion-Based Fusion for Arbitrary Number of Modalities

FlexiD-Fuse (Xu et al., 11 Sep 2025) generalizes to arbitrary $v^i$ 0-way fusion in medical images. It combines:

An ultra-lightweight diffusion backbone (Diffusion Fusion Mamba).
An EM-corrected maximum-likelihood estimator, modeling fusion as inference in a hierarchical latent-variable model.
During inference, an EM loop adapts the fused output to the particular present inputs at every diffusion step, entirely avoiding fixed input assumptions.

3. Practical Workflows and Missing-Data Handling

These frameworks share several modular and operational design patterns for missing modality handling:

Learned Missing-Modal Bank: In PerM-MoE, missing modality embeddings are stably replaced by learned vectors $v^i$ 1 conditioned on the present set, with the same post-processing as genuine data (Burns et al., 8 Sep 2025).
Input-Adaptive Fusion Paths: Both the MedFlexFusion module and FlexCare dynamically modulate the fusion computation graph based on what data is present, using gating signals or masked attention (Zhang et al., 18 Aug 2025, Xu et al., 2024).
No-Imputation Protocols: MedFuse (MuFuse) omits missing tokens entirely, keeping workflows naturally robust and interpretable (Hsieh et al., 12 Nov 2025).
Bayesian EM Correction: FlexiD-Fuse employs a principled EM step at every diffusion iteration, yielding optimal fused representations under arbitrary missingness (Xu et al., 11 Sep 2025).

Workflow integration emphasizes minimal retraining and inference-time adaptability, often recommending precomputation or caching of known-modal embeddings for low-latency deployment.

4. Quantitative Performance and Empirical Impact

MedFlexFusion frameworks achieve empirically superior results under challenging missingness:

Method	Use Case	Most Notable Gains	Source
PerM-MoE	Neuroimaging (AD)	RMSE reductions up to 8.6% (single modality), 7.4% (2-modality), consistent across varying missingness	(Burns et al., 8 Sep 2025)
MedFlexFusion (TGMM)	Cardiac multimodal	2–7% decreases in accuracy/C-index when ablated; consistently best on downstream cardiac tasks	(Zhang et al., 18 Aug 2025)
MedFuse (MuFuse)	EHR time series	Absolute AUPRC gain 0.02–0.03 on ICU mortality and HCC risk, outperforming additive/concat	(Hsieh et al., 12 Nov 2025)
FlexCare	Multitask EHR fusion	AUROC improvements on mortality (0.8823 vs 0.8804) and diagnosis, robust to >85% missingness	(Xu et al., 2024)
FlexiD-Fuse	Med. image (MRI/CT/PET)	Top SSIM/FSIM/Q_NCIE vs. best of 15 SOTA; seamless handling of any input number	(Xu et al., 11 Sep 2025)

These improvements are apparent even when compared to strong baselines—such as SUMMIT, MMF, and Flex-MoE—and are robust across wide ranges of missingness and heterogeneity, generalizing across clinical domains from neurology and cardiology to broad EHR prediction and generative fusion tasks.

5. Model Training, Optimization, and Implementation

Practical deployment of MedFlexFusion models typically leverages standard neural optimization and architectural choices, with modifications to support flexible input:

Standard deep learning optimizers: Adam (or Adafactor), with learning rates in the range $v^i$ 2 to $v^i$ 3.
Architectural depth varies: e.g., 3-layer 3D-CNNs with $v^i$ 4 (Burns et al., 8 Sep 2025), 32-layer Transformers with $v^i$ 5 (Hsieh et al., 12 Nov 2025), or up to $v^i$ 6 final fusion dimension in cardiac analysis (Zhang et al., 18 Aug 2025).
Mixture-of-Experts configurations: Typically $v^i$ 7– $v^i$ 8 experts, top- $v^i$ 9 selection, per-expert MLPs.
Regularization: Load-balancing and expert-specialization losses are common, e.g., as in Shazeer et al. (for MoE), with hyperparameter selection via grid search or Optuna (Burns et al., 8 Sep 2025, Hsieh et al., 12 Nov 2025, Xu et al., 2024).
Batch size and early stopping routines are chosen to match memory and convergence criteria (e.g., batch size 16–32, patience 10–30 epochs).

No explicit imputation routines are required (except for bank-based schemes), and models can be ported to new clinical contexts with modest adaptation—e.g., tuning $E$ 0 or unfreezing the appropriate number of encoder layers.