Machine learning models outperform MELD scores for predicting hepatic encephalopathy mortality

Important points:

• An artificial neural network model better predicted 28-day mortality and MELD, MELD-Na scores in hepatic encephalopathy.
• Independent risk factors – ventilator and acute kidney injury – are important for predicting mortality.

According to the data, a machine learning model using an artificial neural network showed “superior performance” compared to the MELD score for predicting 28-day mortality in patients with hepatic encephalopathy. Department of Gastroenterology, BMC.

“Several studies have shown that outcomes for patients with hepatic encephalopathy (HE) have improved over the past decade, but prognosis and quality of life remain poor.” Zhejiang Zhang, From the Department of Gastroenterology, Tangdu Hospital, China, and a colleague wrote. “Specifically, HE is usually a harbinger of liver decompensation, and its development is usually associated with high morbidity, suggesting the need for liver transplantation.”

Researchers said: Many existing scoring systems have been used to assess the prognosis of patients with cirrhosis, but none have been targeted for HE patients. ”

In a retrospective cohort study, researchers trained and validated a machine learning model to predict mortality in 601 HE patients from an intensive care database medical information mart. Of these patients, 18.64% died within her 28 days.

Zhang and his colleagues developed models using four different machine learning algorithms: artificial neural networks, gradient boosting machines, “random forest” algorithms, and “bagging trees” algorithms. Independent risk factors were acute physiologic score III (APSIII), sepsis-related organ failure assessment (SOFA), international normalized ratio (INR), total bilirubin (TBIL), albumin, blood urea nitrogen (BUN), and acute kidney injury. (AKI) and mechanical ventilation.

The results showed that the neural network model exhibited the largest area under the curve ([AUC] = 0.837; 95% CI, 0.774-0.901) while ‘tree in bag’ showed the lowest discrimination ability (AUC = 0.741; 95% CI, 0.654-0.829). In addition, the neural network model was also ‘well-calibrated’ and improved in predicting 28-day mortality, with an AUC for the previously established MELD and MELD-Na scores of 0.728 (95% CI, 0.677–0.779 ) and 0.711 (95% CI). , 0.658-0.765), respectively.

“We demonstrated that APSIII, SOFA, ventilator, INR, TBIL, albumin, and AKI are important in predicting 28-day mortality,” Zhang and colleagues wrote. “The model in our study performed better than the MELD score or MELD-Na score in predicting 28-day mortality. In the future, real-time prediction of mortality risk in HE patients will be realized. It has the potential to optimize therapy and improve clinical outcomes.”