TinyML: Tools, applications, challenges, and future research directions

Noninvasive Diabetes Detection through Human Breath Using TinyML-Powered E-Nose

“…For the dataset used, the inferiority of KNN and Random Forest in detecting DMI is noticeable, with values of 85% and 90%, approaching false positives. Detecting as many true positives as possible while reducing the number of false positives is crucial [ 16 , 35 , 36 ], as well as understanding the approach of this research in implementing TinyML with TensorFlow Lite on a microcontroller for qualitatively predicting the health status of patients with diabetes mellitus based on ketones related to their BGL, without the need for pre-processing data or the use of the Internet of Things or cloud services [ 53 , 54 , 55 , 56 ]. The most accurate algorithm is XGBoost due to its superior conversion features to a TensorFlow Lite model, detailed in Section 3.4 .…”

“…While the XGBoost model has proven to be the most suitable algorithm for the qualitative detection of low BGL for HI or high levels for DMI, its ability to convert models from different libraries or algorithms to TensorFlow-compatible, and then to TensorFlow Lite models, depends on conversion compatibility [ 55 ]. This poses challenges for algorithms like SVM, KNN, Random Forest, and Decision Tree.…”

“…These algorithms have proven effective in diabetes mellitus detection, as corroborated in previous studies in cloud-based systems or processing results on a computer [ 44 , 51 , 52 ]. While effective, limitations persist, notably with TinyML integration and the explainability of certain algorithms, hindering diagnosis interpretation in case of misdiagnosis [ 53 , 54 , 55 , 56 ]. In contrast, the XGBoost algorithm can be converted to these formats but with the disadvantage of losses in precision and recall metrics, making the process more complicated.…”

“…Consequently, only the top 4 selected characteristics from this process, including carbon monoxide, alcohol, acetone, and benzene, were utilized in the machine learning and deep learning models before training [ 44 , 46 ]. This feature selection process efficiently reduced the number of sensor attributes while retaining the most informative ones [ 35 ], thereby enhancing the performance and computational efficiency of our e-nose-based models post-training and conversion into TensorFlow Lite for the embedded system [ 53 , 55 ].…”

“…Integrating new breath samples into a trained algorithm requires a computer, analyzing differences between samples from sick and healthy individuals [ 44 , 51 , 52 ]. However, there is a solution to this limitation using TinyML, offering edge artificial intelligence (AI) capable of processing data locally, without the need for data transfer, preserving information, reducing latency, energy consumption, and enhancing data privacy/security [ 53 , 54 , 55 , 56 ].…”