Noninvasive evaluation of portal hypertension using a supervised learning technique

Abstract

Portal hypertension (PHT) is a key event in the evolution of different chronic liver diseases and leads to the morbidity and mortality of patients. The traditional reliable PHT evaluation method is a hepatic venous pressure gradient (HVPG) measurement, which is invasive and not always available or acceptable to patients. The HVPG measurement is relatively expensive and depends on the experience of the physician. There are many potential noninvasive methods to predict PHT, of which liver transient elastography is determined to be the most accurate; however, even transient elastography lacks the accuracy to be a perfect noninvasive diagnostic method of PHT. In this research, we are focusing on noninvasive PHT assessment methods that rely on selected best-supervised learning algorithms which use a wide set of noninvasively obtained data, including demographical, clinical, laboratory, instrumental, and transient elastography measurements. In order to build the best performing classification meta-Algorithm, a set of 21 classification algorithms have been tested. The problem was expanded by selecting the best performing clinical attributes using algorithm-specific filtering methods that give the lowest error rate to predict clinically significant PHT. The suggested meta-Algorithm objectively outperforms other methods found in literature and can be a good substitute for invasive PHT evaluation methods.