Search

Acute ischemic stroke prediction from electronic health record time series data holds significant potential for enabling early intervention and reducing long-term disability. LSTMs have been widely used to model clinical sequences such as vital signs and laboratory trends, showing strong performance in stroke-related prediction tasks from 2018–2022. However, their sequential nature limits scalability and long-range dependency modeling in large EHR datasets. Transformers, despite transforming sequence modeling in other domains since 2017, remain underused in stroke prediction compared to LSTMs. Although early healthcare studies suggest potential benefits of attention-based models, robust validation in acute ischemic stroke contexts is still limited. Transformers offer advantages in parallel processing, long-range dependency modeling, and interpretability, but require more data and computational resources. They are likely to complement rather than replace LSTMs, with hybrid architectures providing a balanced solution for clinical time series analysis. Key themes include long-range dependency capture, parallel computation, interpretability, and data efficiency trade-offs between LSTMs and transformers. Hybrid LSTM–transformer models may offer improved performance and practicality for stroke prediction, with model selection depending on data scale and clinical constraints. Further benchmarking is needed to determine when transformers or hybrid models outperform LSTMs, guiding the development of more effective stroke prediction systems.

Chronic kidney disease (CKD) affects 10–15% of adults worldwide and often progresses silently toward kidney failure requiring dialysis or transplantation. Monitoring longitudinal markers such as estimated glomerular filtration rate, creatinine, and albuminuria is essential for early intervention and delaying disease progression. However, current predictive models typically rely on static or isolated clinical features, limiting their ability to capture dynamic interactions between laboratory trends, medications, and comorbidities, which leads to incomplete risk assessment. To address this limitation, a conceptual framework based on a graph convolutional network with attention mechanisms is proposed to integrate longitudinal laboratory data, medication networks, and comorbidity structures for CKD progression prediction. Patient records from 2017–2023 are represented as a heterogeneous graph, where nodes include laboratory values, drugs, and diagnoses, and edges encode clinical and pharmacological relationships. Graph convolutional layers capture relational patterns, while attention mechanisms highlight the most clinically relevant interactions, enabling more informative patient-level representations for risk prediction across CKD stages. This approach improves interpretability by revealing which laboratory trends, medications, and comorbidities most influence predicted outcomes, aligning model behavior with clinical nephrology knowledge. Overall, the framework provides a unified and scalable strategy for more accurate and interpretable CKD progression risk prediction by leveraging relational and temporal data structures that traditional models fail to exploit.

Immunotherapy with immune checkpoint inhibitors is a standard treatment for advanced non-small cell lung cancer (NSCLC), with durable responses in selected patients. Whole-slide histopathology images provide morphological and immune microenvironment information, while genomic expression data capture pathway activity and resistance mechanisms. Single-modality approaches based on either histopathology or genomics fail to capture complementary tumor information, limiting accurate stratification of responders and non-responders and leading to suboptimal treatment selection. We propose a multimodal fusion network that integrates whole-slide histopathology images and genomic expression data to predict immunotherapy response in NSCLC. Separate encoders process each modality, followed by cross-attention for joint representation learning in an end-to-end framework. The system includes a multiple instance learning-based WSI module, a gene expression encoder with attention over gene sets, and a cross-attention fusion module. The model outputs a binary or probabilistic prediction of treatment response using paired slide and genomic data. The model captures complementary morphological and molecular signals, linking immune infiltration patterns with transcriptomic activity. Attention mechanisms enhance interpretability by highlighting key tissue regions and gene pathways, while also improving robustness to partial modality missingness. This multimodal framework improves NSCLC immunotherapy response prediction by integrating histopathology and genomic data, offering a step toward more precise patient stratification in precision oncology.