Investigation of Non-Invasive Attributes for Classification of Patients with Portal Hypertension

Abstract

BACKGROUND: Portal hypertension (PHT) is the key indicator of evolving chronic liver diseases. Standard clinical diagnostic method is based on invasive and inconvenient procedure. OBJECTIVE: The main goal was to create an objective machine learning method for evaluating PHT by selecting of most informative attributes derived only from noninvasive investigations. METHOD: A proposed meta-algorithm selects five best performing standard classification algorithms by AUC parameter from five typical groups. The best performing noninvasive attributes were selected by testing three ranking methods: correlation, relief and relative frequency of occurrence (RFO). Invasively measured hepatic venous pressure gradient (HVPG) served as class attribute: HVPG < 10mmHg and HVPG $\pmb{\geq 10}$ mmHg. The missing values (MVs) in data, were imputed by using regression based Iterative Robust Model-Based Imputation (IRMI) algorithm. RESULTS: The number of selected most informative attributes was 4 out of total 24 by RFO method. The meta-algorithm resulted with AUC = 0.97 and classification accuracy of 90.22%. CONCLUSIONS: The RFO method allows ranking and selecting most informative attributes objectively. Meta-algorithm objectively outperforms other noninvasive methods and can be a good candidate to substitute invasive PHT evaluation methods because.