Dynamic neural networks in biomedical applications: A review from embedded systems to AI-driven interventions

abstract

Dynamic Neural Networks are redefining biomedical intelligent systems by bridging mathematical control theory and clinical translation. Rooted in differential formulations originally developed for nonlinear system identification, Dynamic Neural Networks extend beyond static learning to capture continuous-time feedback, memory, and adaptation¿core properties of living physiological systems. This review unifies theoretical and applied developments of Dynamic Neural Networks across biomedical domains, highlighting their capacity to model, identify, and control complex biological dynamics. Using the PRISMA methodology, publications from 2015¿2025 were systematically analyzed to map the evolution of recurrent, differential, and hybrid neural architectures across four major application areas: physiological signal monitoring and interpretation, intelligent control systems in therapeutic devices, drug delivery and pharmacological regulation, and personalized medicine with predictive diagnostics. The reviewed works demonstrate how Dynamic Neural Networks-based frameworks enable context-aware interpretation of biosignals, adaptive control in prosthetics and exoskeletons, and data-driven precision in clinical decision support. Notably, a differential neural identifier for fetal ECG monitoring is presented as a case study, exemplifying how continuous-time learning can reconstruct hidden cardiac dynamics from single-channel abdominal recordings¿achieving accurate morphology preservation and noise rejection with demonstrated suitability for embedded or resource-constrained fetal monitoring scenarios. Across applications, Dynamic Neural Networks consistently exhibit advantages over static architectures in terms of adaptability, feedback integration, and temporal modeling capability, which are critical properties for real-time biomedical systems. While fully deployed embedded implementations remain limited to selected studies, the reviewed evidence indicates strong algorithmic suitability and growing feasibility for embedded and edge-AI realization. The findings confirm that Dynamic Neural Network¿based identifiers, controllers, and predictors form the computational foundation for next-generation intelligent healthcare, advancing the convergence of theoretical dynamic modeling, embedded artificial intelligence, and adaptive therapeutic intervention. © 2026 Elsevier B.V.