An Adaptive Recursive Attention Network for Medical Equipment Monitoring via IoT-Integrated AI Systems

Abstract

In the context of challenges faced by remote monitoring and maintenance of medical equipment, this study proposes an adaptive recursive attention network (ARAN) model to improve the accuracy of equipment fault diagnosis and remaining useful life prediction. The ARAN architecture integrates a recurrent neural network framework with an attention mechanism that adaptively emphasizes relevant time steps in the input sequence. The model was trained and evaluated on the Medical Device Operation Dataset (MDO-Dataset), which comprises time-series operational data from over 5,000 medical devices across multiple large hospitals over five years. ARAN was benchmarked against three baseline models: Recurrent Neural Network (RNN), Long Short-Term Memory (LSTM), and Gated Recurrent Unit (GRU). Quantitative results show that ARAN achieved a fault diagnosis accuracy of 91.25%, outperforming the baseline average of 70%, with category 4 reaching 95%. For remaining useful life prediction, ARAN yielded a mean absolute error (MAE) of 53.25 hours, substantially lower than the baseline MAE of 100 hours. Additionally, ARAN demonstrated superior performance in recall rate, F1 score, false alarm rate, and noise robustness, with faster convergence speed and stronger generalization ability on unseen device types. This research presents an effective fusion of IoT and AI technologies for intelligent medical equipment management.