A Fast and Reliable Transformer-Based TabPFN Model for Liver Disease Diagnosis

Early and accurate diagnosis of liver disease is the key to timely treatment and improved patient outcomes. Traditional machine learning models often utilize manually constructed features and struggle to extrapolate to high-dimensional, complex tabular data. This work presents an original approach to predicting liver disease using TabPFN (Prior-Data Fitted Network for Tabular Data), a pre-trained transformer-based model on artificial tabular problems. This model incorporates one-dimensional feature attention and one-dimensional sample attention techniques into the model, and subsequently a multilayer perceptron classifier. Experimental results on the Indian Liver Patient Dataset demonstrate better performance than existing approaches, with an accuracy of 90.36%, an F1-score of 90.35%, a precision of 90.99%, and a recall of 90.36%. These results show how the suggested architecture model effectively captures relevant patterns in tabular information for liver disease identification.