A Multimodal Approach to Liver Cirrhosis Detection Using Ensemble Machine Learning Algorithms

Abstract

Liver cirrhosis is a severe and progressive disease marked by the deterioration of liver cells and the formation of fibrous scar tissue. This condition often presents late in its course, making timely and accurate diagnosis crucial for effective management and treatment. This thesis introduces ""HappyLiver,"" an advanced liver cirrhosis detection system that utilizes a hybrid machine learning model to improve diagnostic accuracy significantly. The ""HappyLiver"" system employs a multimodal data fusion approach that integrates three distinct types of data: biochemical test results, synthetic symptom data, and demographic information. The biochemical data includes liver function tests such as mean corpuscular volume, alkaline phosphatase levels, and gamma-glutamyl transpeptidase, among others. The symptom data, artificially generated for this study, simulates common cirrhosis-related symptoms such as fatigue, nausea, and fever. Demographic data, including age and gender, are also considered, given their relevance to disease prevalence and manifestation. The core of the detection system is a sophisticated machine learning algorithm incorporating ensemble methods, specifically a stacking classifier that combines the predictive power of random forests and gradient boosting models, with logistic regression as a meta-classifier. This model structure was chosen based on its ability to handle diverse data inputs and provide robust, generalized predictions that outperform traditional single-algorithm approaches. This thesis not only demonstrates the effectiveness of using multimodal data and ensemble machine learning techniques in medical diagnosis but also highlights the potential for such systems to be adapted for other complex diseases.