Comparative Evaluation of Deep Learning Models for Exhaled Breath-Based Liver Disease Prediction

Liver diseases are on the rise due to factors such as excessive alcohol intake, poor dietary habits, and sedentary lifestyles, along with metabolic disorders like obesity and diabetes. For instance, non-alcoholic fatty liver disease (NAFLD) affects nearly a quarter of the global population, and liver cirrhosis causes over 1.32 million deaths annually, as reported by the WHO. Traditional diagnostic methods, including liver biopsies and imaging, are invasive, expensive, and impractical for large-scale or routine screening. This creates a need for non-invasive, cost-effective, and reliable diagnostic alternatives. One such approach is exhaled breath analysis, which utilizes volatile organic compounds (VOCs) to detect liver dysfunction. However, analyzing this data effectively requires advanced computational techniques capable of handling complex, high-dimensional sequences. This study explores the performance of various deep learning models—LSTM, BiLSTM, GRU, and 1D CNN—in predicting liver disease. Each architecture has distinct advantages: LSTM and BiLSTM efficiently capture long-term dependencies in sequential data, GRU provides computational efficiency while maintaining accuracy, and 1D CNNs are effective in feature extraction from raw input. By evaluating these models based on accuracy, sensitivity, and computational efficiency, this research aims to identify their respective strengths and limitations in liver disease classification. The findings will aid in selecting the most suitable model for real-time applications, ensuring both high predictive accuracy and optimal computational performance. Ultimately, this work contributes to the advancement of non-invasive diagnostic tools, facilitating early and precise liver disease detection while minimizing the need for invasive procedures and enhancing patient care.