A Framework for Enhancing Shelf Life Prediction of Perishable Foods Using Machine Learning

Thesis Abstract

Perishable foods are highly susceptible to spoilage due to microbial activity, enzymatic reactions, and environmental factors, leading to significant economic losses and health risks, thereby necessitating accurate and reliable shelf life prediction models. Despite advancements in food preservation techniques, existing methods for shelf life estimation often rely on empirical approaches, which lack adaptability and precision across diverse food matrices and storage conditions. This study aims to develop a comprehensive machine learning-based framework to enhance the accuracy and robustness of shelf life prediction for perishable foods, thereby supporting supply chain optimization, reducing waste, and ensuring consumer safety. The specific objectives include (1) identifying key spoilage indicators and the corresponding environmental parameters influencing shelf life; (2) designing data collection protocols for relevant physicochemical, microbiological, and sensory attributes; (3) evaluating and comparing the predictive performance of various machine learning algorithms such as random forests, support vector machines, and artificial neural networks; and (4) formulating an integrative predictive framework tailored to perishable food products. The research adopts a mixed-methods approach, combining quantitative modeling with qualitative analysis of the influence of external variables on spoilage patterns. The population consists of perishable food samples—primarily fresh-cut fruits, dairy products, and seafood—collected from three different commercial outlets over a period of six months. A stratified random sampling technique is employed, resulting in a sample size of 1,200 units to ensure representativeness and statistical validity. Data collection involves the use of multi-parameter sensors for measuring temperature, humidity, and gas composition; microbiological assays to quantify spoilage microorganisms; physicochemical analyses such as pH, water activity, and oxygen levels; and sensory evaluation conducted by trained panels. The instruments used include portable gas analyzers, microbiological culture kits, pH meters, and sensory scoring sheets. Data are processed using descriptive statistics to understand baseline variations, while advanced analytical techniques—such as principal component analysis (PCA)—are used for feature reduction. Machine learning algorithms are trained and validated through k-fold cross-validation, with performance evaluated based on metrics like root mean square error (RMSE), accuracy, precision, recall, and F1-score. Furthermore, the study incorporates theoretical underpinnings from the Information Processing Theory and the Diffusion of Innovations Theory, to understand how predictive insights influence decision-making processes within the food supply chain. The anticipated findings include the identification of critical spoilage predictors, improved predictive accuracy over traditional models, and the development of an adaptable framework that can accommodate various food types and storage conditions. The framework is expected to outperform existing empirical models by providing dynamic, real-time shelf life estimations that account for environmental fluctuations. The contribution of this research lies in bridging the gap between food science and data science by establishing an innovative, scalable, and practical model for shelf life prediction. It provides a foundation for food professionals, supply chain managers, and regulatory agencies to implement evidence-based preservation strategies, optimize logistics, and minimize product wastage. The study concludes that machine learning techniques, when integrated within a structured analytical framework, significantly enhance the reliability of shelf life estimates and support sustainable food systems. Based on the findings, recommendations include the adoption of real-time sensor technologies, continuous model updating with new data, and further research into integrating predictive frameworks with blockchain for traceability. Future studies should explore the application of deep learning models and expand the scope to include emerging perishable products with rapid spoilage dynamics.