Towards Reliable Multimodal Clinical Decision Support: From Quality-Aware Fusion to Patient Digital Twins

1 Introduction and Research Vision

Deep learning models for clinical decision support increasingly rely on multimodal inputs, integrating electronic health records (EHR), medical imaging, and continuous vitals. While theoretical performance is strong, real-world deployment frequently fails. Mainstream multimodal fusion and MoE gating mechanisms [Shazeer et al., 2017; Zhou et al., 2022] operate under the assumption that all available modalities provide uniform benefits. In practice, clinical data suffers from severe structural missingness, sensor noise, and distribution shifts. When weakly informative or corrupted modalities are allowed to perturb the routing process, the network experiences routing collapse and miscalibrated confidence.

My doctoral research addresses the gap between idealized architectures and messy clinical reality. Instead of relying on generative imputation [Yoon et al., 2018], which risks hallucinating clinical features, my work shifts the paradigm towards uncertainty-guided abstention and dynamic structural adaptation. The overarching vision of my PhD is to develop robust inference frameworks that can gracefully handle missingness and temporal shifts, ultimately serving as the computational foundation for predictive Patient Digital Twins.

1.1 Current Work: Quality-Aware Multimodal Routing

To prevent noisy modalities from degrading overall predictions, I developed QA-MoE, a modular architecture that explicitly decouples reliability estimation from the routing process. As illustrated in Figure 1, traditional MoEs mix feature extraction and routing, making them vulnerable to corrupted inputs. In QA-MoE, an Evidential Quality Scorer first quantifies the epistemic uncertainty of each modality. This score guides a Stability-Enhanced Subset Selector to dynamically filter out noisy signals before they reach the experts. For cases of severe structural missingness, I introduced a Ternary Expert Aggregation branch that acts as a fail-safe. This mechanism ensures that the model maintains sparse, interpretable, and stable fusion even when critical modalities drop out unexpectedly.

1.2 Ongoing and Future Directions: Temporal and Topological Dynamics

Building on the static robustness of QA-MoE, my current focus expands to longitudinal patient trajectories and continuous physiological tracking. To address the inherent irregularity and sparsity of clinical events, I am developing PathFlowEM, a framework that integrates Expectation-Maximization to explicitly infer hidden variable distributions across complex, unevenly sampled clinical time-series data. In parallel, to scale robust sequence modeling to high-frequency clinical streams, I have designed Phase-Mamba, a linear state-space architecture for continuous physiology forecasting. Evaluated on the MIMIC-IV and PhysioNet 2019 datasets for cross-center validation, Phase-Mamba regulates information flow by applying a rigorous filtering threshold to the normalized Shannon Entropy of the system dynamics rather than a standard gating coefficient, thereby effectively isolating unpredictable physiological noise. Concurrently, to handle the high-dimensional structural relationships within multi-omics and patient similarity networks, my future work involves UDTF. By applying topological data analysis [Hensel et al., 2021] to track multi-scale persistent homological features, UDTF aims to gate features based on evolving graph structures rather than treating patient data as isolated snapshots.

2 Evaluation Strategy and Broader Impacts

Evaluating clinical models requires metrics that go beyond basic accuracy, prioritizing calibration and computational efficiency under duress.

2.1 Empirical Validation on Clinical Benchmarks

My frameworks are tested on heavily noisy, real-world datasets. For instance, QA-MoE was evaluated on the ADNI dataset for Alzheimer's staging (combining imaging and cognitive scores) and the MIMIC-IV dataset for Length-of-Stay prediction (dense EHR time-series). The results demonstrate that QA-MoE not only outperforms strong multimodal baselines in predictive accuracy but also maintains low Expected Calibration Error (ECE) under simulated missingness ranging from 10% to 50%. Furthermore, bypassing noisy experts on the fly significantly cuts down unnecessary computation overhead. Specifically, we compare QA-MoE against state-ofthe-art multimodal fusion baselines, including standard MoE routing, multimodal gating architectures, and traditional nonMoE models. Under severe simulated missingness, QA-MoE reduces the ECE while simultaneously achieving the highest accuracy, F1, and AUC scores.

2.2 Societal Impact and Ethical Considerations

As the healthcare sector moves towards continuous patient monitoring, the ability to synthesize multimodal streams without catastrophic failure from missing data becomes a prerequisite for creating reliable Patient Digital Twins. However, deploying such models introduces strict ethical constraints. In medical AI, an overconfident incorrect prediction is far more dangerous than an abstention. By explicitly modeling epistemic uncertainty at the modality level, my architecture prevents the system from making high-confidence guesses on severely incomplete patient profiles. Future iterations of this work will also incorporate privacy-preserving constraints into the topological graphs to ensure that learning from patient similarity networks does not compromise individual confidentiality or amplify historical biases.