Expanding Applications of TinyML in Versatile Assistive Devices: From Navigation Assistance to Health Monitoring System Using Optimized NASNet-XGBoost Transfer Learning

• Published in: IEEE access 2024, Vol.12, p.168328-168338
• Main Authors: Ponnada, Sreenu, Kuan Tak, Tan, Kshirsagar, Pravin R., Srinivasa Rao, P., Dayal, Abhinav
• Format: Article
• Language: English
• Subjects: Accuracy; Adaptation models; Brain modeling; Computational modeling; Data models; Data processing; Data storage; Diseases; EWT-SAS; Feature extraction; Health care; Human factors; Learning; Medical services; Monitoring; NASNet-XGBoost; Performance measurement; Portable equipment; Signal processing; Telemedicine; Tiny machine learning; TinyML; tyrannosaurus optimization; various disease prediction; wavelet soft threshold denoising; Wavelet transforms
• ISSN: 2169-3536; 2169-3536
• DOI: 10.1109/ACCESS.2024.3496791

The healthcare sector receives a considerable amount of unprocessed data from wearable and portable devices. However, traditional cloud-based models used to handle this type of data can pose risks such as exposing sensitive patient data to a network environment and increasing latency due to data storage and time consumption. In response to these issues, an alternative strategy named TinyML was put forth by the scientific community to foster safe and autonomous models capable of gathering and processing data without the need for network exposure. To address these concerns more effectively, we put forward a versatile TinyML model specifically tailored for health monitoring systems. This model uses enhanced NASNet-XGBoost-based transfer learning methodology to adapt to various types of diseases using appropriate healthcare data. The process begins by pre-processing raw health signals collected from patients using a combination of a wavelet soft threshold method and empirical wavelet transform-spectrum adaptive segmentation (EWT-SAS). A Tyrannosaurus-based NASNet model is then used to extract meaningful features from these pre-processed signals. Once the features have been extracted, they are entered into an XGBoost model which can accurately predict various types of diseases based on the given data. This particular approach achieved an impressive 95.4% average accuracy, 93.6% positive predictive value, 94.7% hit rate, and 96.3% selectivity. When compared to traditional models, this TinyML model shows a significant improvement in performance metrics, demonstrating increased effectiveness and accuracy in predicting various diseases.