Real-time evaluation of seafood freshness is essential for ensuring food quality and safety, as approximately 35% of seafood is lost or wasted due to quality deterioration. This study introduces a portable, real-time electronic nose (e-nose) system designed to assess freshness and predict remaining shelf life using machine learning (ML) algorithms. The hardware utilizes a gas sensor array connected to an ESP Devkit-C S3 microcontroller to detect volatile organic compounds (VOCs) emitted during spoilage. From an initial pool of eight sensors, four were selected based on Predictive Power Score (PPS) analysis for their high sensitivity to spoilage indicators: MQ137 (ammonia), MQ2 (alcohol), MQ136 (hydrogen sulphide), and MQ9 (methane). The core of the system is the E-Nose-Based Seafood Freshness (ESF) dataset, which contains 336,997 data points from tuna, salmon, cod, shrimp, and crab.

To address class imbalance, the dataset was processed using the Synthetic Minority Over-sampling Technique (SMOTE), resulting in 457,954 samples for robust training. Various ML models, including Random Forest, KNN, Naive Bayes, Decision Tree, AdaBoost, Gradient Boosting, and XGBoost, were evaluated for classification and regression tasks. XGBoost emerged as the superior algorithm, achieving 100% precision, recall, and accuracy in classification, with a detection time of under 1 millisecond and a memory footprint of approximately 4 kB. For regression tasks, an RMSE of 0.08. These results confirm the proposed system as a lightweight, scalable, and highly accurate solution for the seafood supply chain, providing a non-invasive and cost-effective alternative to traditional chemical or microbiological methods.