PAMF: Prior-Aware Multimodal Fusion for Incomplete Time Series Data

PAMF is a new machine learning framework that addresses incomplete multimodal time series data in healthcare by distinguishing between two types of missing data patterns and coupling imputation with downstream prediction tasks. The method uses flow matching with type-specific priors and weight sharing to achieve superior performance on healthcare benchmarks compared to existing approaches.

PAMF represents a meaningful advancement in handling incomplete medical data, a persistent real-world challenge in healthcare informatics. The framework tackles two structurally different missingness patterns—within-modality gaps (missing values within an observed signal) and modality-level gaps (entire channels unavailable)—recognizing that treating all missing data uniformly degrades performance. Current methods rely on implicit masking or generic embeddings without leveraging the unique characteristics of each missingness type.

The innovation stems from healthcare's practical constraints where patient monitoring equipment frequently fails. ECG electrodes detach, respiratory monitors malfunction during overnight monitoring, and data collection isn't always complete. Existing imputation techniques exist in isolation from prediction tasks, creating a disconnect where the imputed values optimize for reconstruction rather than downstream diagnostic utility. PAMF closes this gap through weight-shared encoders that allow prediction tasks to guide imputation toward clinically relevant representations.

The technical approach initializes flow-matching source states with type-specific structural priors, enabling the model to handle different missingness patterns with tailored assumptions. Architectural matching between imputation and classification components with shared weights creates bidirectional information flow—predictions inform what aspects of missing data matter most for diagnosis. Experimental validation across multiple healthcare benchmarks demonstrates consistent performance gains over baseline approaches across diverse missing data scenarios.

For healthcare AI developers and medical informatics teams, this work provides a principled framework for building robust diagnostic systems that function reliably despite inevitable data gaps. The research suggests that next-generation clinical decision support systems should explicitly model missingness heterogeneity rather than applying uniform masking strategies.

• →PAMF distinguishes between within-modality and modality-level missingness using type-specific structural priors rather than uniform treatment
• →Weight sharing between imputation and prediction components allows downstream tasks to guide which missing values matter most for diagnosis
• →Flow matching with task-coupled architecture achieves superior performance across multiple healthcare time-series benchmarks
• →The framework addresses a practical healthcare challenge where equipment failure and incomplete monitoring are common operational realities
• →Bidirectional information flow between imputation and prediction optimizes for clinically relevant representations rather than reconstruction accuracy alone