Longitudinal health records provide a rich temporal view of patient trajectories, enabling early identification of disease progression before overt clinical deterioration occurs. As healthcare systems increasingly rely on continuous data streams from electronic health records, wearable sensors, and remote monitoring platforms, robust temporal pattern recognition methods have become critical for proactive clinical decision-making. However, early warning systems in healthcare must operate under conditions of sparse labels, heterogeneous data sources, and evolving disease dynamics, making unsupervised anomaly detection particularly well suited to this domain. This work examines temporal pattern recognition and unsupervised anomaly detection frameworks for early detection of disease progression in longitudinal health records. From a broad perspective, the study situates anomaly detection within predictive population health, chronic disease management, and preventive care, emphasizing its role in identifying subtle deviations from personalized baselines rather than population-wide thresholds. The discussion then narrows to temporal modeling techniques, including sequence-aware statistical models, representation learning approaches, recurrent and attention-based architectures, and time-series clustering methods that capture latent disease dynamics without reliance on labeled progression events. Key methodological challenges are addressed, including handling irregular sampling, missing data, patient heterogeneity, and non-stationary clinical processes. The abstract further highlights evaluation strategies that prioritize early detection sensitivity, temporal consistency, and clinical relevance, rather than retrospective classification accuracy alone. Particular attention is given to minimizing false alarms and supporting clinician interpretability to ensure practical adoption in real-world settings. The study concludes by identifying future research directions, including integration of multimodal longitudinal data, adaptive personalization of anomaly thresholds, and alignment with explainability and safety requirements for clinical deployment. Temporal unsupervised anomaly detection is positioned as a foundational capability for scalable, early-warning disease surveillance and individualized care pathways.