Lightweight Deep Learning Models for Real-Time Anomaly Detection in Critical Hospital IoT Environments

Abstract

The rapid proliferation of Internet of Things (IoT) devices within modern hospital environments has significantly reshaped clinical workflows, biomedical device coordination, and real-time patient monitoring. As hospitals increasingly transition toward interconnected smart infrastructures, the reliability and security of these devices become central determinants of patient safety and operational stability. Real-time anomaly detection plays an essential role in mitigating risks associated with device malfunction, abnormal physiological readings, environmental fluctuations, and potential cybersecurity threats. However, conventional deep learning techniques often exceed the computational capacity of embedded medical IoT hardware, which typically operates with tight constraints on memory, processing power, and energy consumption. 

This article explores lightweight deep learning architectures that achieve real-time inference on edge-deployed medical IoT devices without compromising detection accuracy. The study evaluates MobileNet autoencoders, micro-temporal convolutional networks (micro-TCNs), and compressed LSTM variants—models chosen for their capacity to scale down while retaining expressive temporal modeling. Using representative hospital IoT datasets across four device categories—vital-sign monitors, infusion pumps, RFID-based asset trackers, and environmental sensors—we conduct extensive experiments on latency, energy consumption, detection performance, and robustness to noise and device variability. 

The results demonstrate that lightweight architectures deliver competitive detection accuracy with sub-second latency, enabling autonomous, on-device anomaly detection without reliance on cloud connectivity. This work offers a systematic, 5000-word examination of computation-efficient neural models, their architectural considerations, optimization techniques such as quantization and pruning, and their applicability to critical hospital IoT environments. The findings contribute directly to the design of reliable, secure, and scalable AI-driven hospital infrastructures.