This study presents a noninvasive method for detecting diabetes mellitus using an e-nose equipped with Metal Oxide Semiconductor (MOS) sensors and powered by TinyML for real-time classification. The system analyzes volatile organic compounds (VOCs) in exhaled human breath, with acetone serving as a key biomarker for diabetes. The e-nose system was tested on 44 participants, including 22 healthy individuals and 22 diagnosed with diabetes (T1DM and T2DM). The study employed machine learning algorithms, including XGBoost, Deep Neural Networks (DNN), and 1D-CNN, to classify breath samples. XGBoost achieved 95% detection accuracy, while DNN and 1D-CNN achieved 94.44% accuracy. These results demonstrate the effectiveness of integrating e-noses with TinyML in embedded systems for noninvasive diabetes detection. The e-nose system includes six MOS sensors, a dehumidifier, and a DHT-22 sensor for humidity and temperature monitoring. Breath samples were collected using Tedlar bags, and preprocessing steps such as Discrete Wavelet Transform (DWT) and Z-score normalization were applied to enhance data quality. Feature selection focused on key VOCs, including acetone, carbon monoxide, alcohol, and benzene. The system was implemented on an Arduino Nano 33 BLE Sense microcontroller, with models converted to TensorFlow Lite for efficient on-device processing. The study highlights the potential of TinyML and embedded systems for real-time, noninvasive diabetes detection. However, challenges such as sensor sensitivity to humidity, the need for dehumidification, and the limitations of small sample sizes were identified. Future research should focus on improving model interpretability, expanding participant diversity, and optimizing sensor integration for broader applicability. The integration of TinyML with e-noses offers a promising solution for early and accurate diabetes detection, reducing the need for invasive blood tests and enhancing healthcare accessibility.This study presents a noninvasive method for detecting diabetes mellitus using an e-nose equipped with Metal Oxide Semiconductor (MOS) sensors and powered by TinyML for real-time classification. The system analyzes volatile organic compounds (VOCs) in exhaled human breath, with acetone serving as a key biomarker for diabetes. The e-nose system was tested on 44 participants, including 22 healthy individuals and 22 diagnosed with diabetes (T1DM and T2DM). The study employed machine learning algorithms, including XGBoost, Deep Neural Networks (DNN), and 1D-CNN, to classify breath samples. XGBoost achieved 95% detection accuracy, while DNN and 1D-CNN achieved 94.44% accuracy. These results demonstrate the effectiveness of integrating e-noses with TinyML in embedded systems for noninvasive diabetes detection. The e-nose system includes six MOS sensors, a dehumidifier, and a DHT-22 sensor for humidity and temperature monitoring. Breath samples were collected using Tedlar bags, and preprocessing steps such as Discrete Wavelet Transform (DWT) and Z-score normalization were applied to enhance data quality. Feature selection focused on key VOCs, including acetone, carbon monoxide, alcohol, and benzene. The system was implemented on an Arduino Nano 33 BLE Sense microcontroller, with models converted to TensorFlow Lite for efficient on-device processing. The study highlights the potential of TinyML and embedded systems for real-time, noninvasive diabetes detection. However, challenges such as sensor sensitivity to humidity, the need for dehumidification, and the limitations of small sample sizes were identified. Future research should focus on improving model interpretability, expanding participant diversity, and optimizing sensor integration for broader applicability. The integration of TinyML with e-noses offers a promising solution for early and accurate diabetes detection, reducing the need for invasive blood tests and enhancing healthcare accessibility.