Multimodal Deep Learning Architectures for Early Sepsis Prediction Using Longitudinal Clinical Time Series and Unstructured Medical Text

Early detection of sepsis remains a persistent challenge in acute and critical care due to the heterogeneous, temporal, and multimodal nature of clinical data preceding disease onset. Traditional rule-based scores and unimodal predictive models often fail to provide sufficient lead time for effective intervention, as they rely on static thresholds or limited representations of patient state. This study proposes a multimodal deep learning framework for early sepsis prediction that jointly models longitudinal clinical time series and unstructured medical text. The architecture integrates transformer-based temporal encoders for physiological signals and laboratory trends with domain-adapted language models for clinical narratives, coupled through a cross-modal attention fusion mechanism that supports asynchronous and partially observed data. The model is evaluated across multiple clinically relevant prediction horizons, with performance assessed using AUROC, AUPRC, and lead-time gain metrics. Results demonstrate that the multimodal approach consistently outperforms traditional risk scores, classical machine learning models, and unimodal deep learning baselines, particularly at longer lead times where early signals are sparse. Ablation and robustness analyses confirm the critical contribution of clinical text and cross-modal attention to early detection performance and stability under missing or delayed data conditions. Interpretability analyses further show that model predictions align with established clinical reasoning, highlighting salient physiological trends and meaningful narrative cues. This work illustrates the potential of multimodal deep learning to enable proactive sepsis management by delivering earlier, interpretable, and clinically actionable risk assessments. The proposed framework provides a foundation for next-generation clinical decision support systems that move beyond reactive detection toward anticipatory care.