Ruizhe Ma

Deep learning has transformed healthcare by enabling the analysis of complex, high-dimensional, and heterogeneous data. However, traditional unimodal approaches often fail to capture the multifaceted nature of human health, as patient information is inherently distributed across multiple data modalities. Multimodal machine learning (MML) has therefore emerged as a framework for integrating complementary sources such as medical images, clinical text, electronic health records (EHRs), and physiological signals to support more comprehensive modeling of health and disease. This narrative review provides a structured overview of MML in healthcare, focusing on representative data modalities, fusion strategies, advanced architectures, and clinically relevant design trade-offs. In particular, we distinguish between stage-based fusion strategies, which determine when modalities are combined, and feature integration mechanisms, which define how modality representations are merged. We synthesize applications across major domains, including brain disorders, cancer prediction, chest-related conditions, and skin diseases, highlighting both the potential benefits and the limitations of multimodal approaches. We further discuss key challenges related to data heterogeneity, cross-modal alignment, missing modalities, and the complexity of effective fusion design, along with broader issues in clinical translation. Finally, we outline future directions centered on foundation models, causal reasoning, privacy-preserving learning, and integrated healthcare data infrastructures.

The social link prediction poses a fundamental challenge in social network analysis, aiming to forecast missing interactions among users. Given the dynamic evolution mechanism of social networks, prevailing efforts have introduced embedding-based approaches to address temporal link prediction in dynamic social networks. However, these approaches often struggle to handle explainability, multirelations, and multihop relation prediction simultaneously. To overcome these limitations, we present an innovative multihop temporal social link prediction model based on temporal logic embedding (TLE), which leverages temporal knowledge graphs and logic rules. First, inspired by the temporal knowledge graph, we construct temporal social knowledge graphs (TSKGs) to model dynamic social networks. Then, we incorporate the orthogonal transformation matrix into the graph neural networks (GNNs), thereby facilitating the learning of time-aware relation representations. Furthermore, we define temporal social random walks from the TSKG to generate temporal social rules. Subsequently, TLE employs time-aware relation embedding to calculate the confidence associated with each rule. Finally, TLE combines the confidence score and time difference to obtain the final link prediction, providing explainability for multihop link predictions. The experiments carried out on four datasets indicate the superiority of TLE in the social link prediction within dynamic social networks.

Time series forecasting is widely used in finance, meteorology, and industrial systems. Although existing methods have made progress in modeling trend and periodicity, they still face challenges when handling real-world data characterized by fuzziness and nonlinear variations. To address this, we propose FPSO-Time, a fuzzy time series forecasting model based on Fuzzy Particle Swarm Optimization (FPSO) and reprogrammed large language models (LLMs). The proposed method employs the FPSO algorithm to optimize the fuzzy universe partitioning of time series data and incorporates triangular membership functions for fuzzification, improving the accuracy and stability of fuzzy logical relationship modeling. Meanwhile, we leverage a GPT-2 based LLM to predict fuzzy membership sequences, with final forecasting results obtained through defuzzification. Furthermore, we design a modular framework that offers strong generalizability, enabling flexible integration with various time series forecasting models and enhancing prediction performance in noisy environments. Experimental results demonstrate that FPSO-Time outperforms existing state-of-the-art methods on multiple public datasets, validating its effectiveness and practical utility.