Artificial neural network model for preoperative prediction of severe liver failure after hemihepatectomy in patients with hepatocellular carcinoma

Abstract

Background

Posthepatectomy liver failure is a worrisome complication after major hepatectomy for hepatocellular carcinoma and is the leading cause of postoperative mortality. Recommendations for hepatectomy for hepatocellular carcinoma are based on the risk of severe posthepatectomy liver failure, and accurately predicting posthepatectomy liver failure risk before undertaking major hepatectomy is of great significance. Thus, herein, we aimed to establish and validate an artificial neural network model to predict severe posthepatectomy liver failure in patients with hepatocellular carcinoma who underwent hemihepatectomy.

Methods

Three hundred and fifty-three patients who underwent hemihepatectomy for hepatocellular carcinoma were included. We randomly divided the patients into a development set (n = 265, 75%) and a validation set (n = 88, 25%). Multivariate logistic analysis facilitated identification of independent variables that we incorporated into the artificial neural network model to predict severe posthepatectomy liver failure in the development set and then verified in the validation set.

Results

The morbidity of patients with severe posthepatectomy liver failure in the development and validation sets was 24.9% and 23.9%, respectively. Multivariate analysis revealed that platelet count, prothrombin time, total bilirubin, aspartate aminotransferase, and standardized future liver remnant were all significant predictors of severe posthepatectomy liver failure. Incorporating these factors, the artificial neural network model showed satisfactory area under the receiver operating characteristic curve for the development set of 0.880 (95% confidence interval, 0.836–0.925) and for the validation set of 0.876 (95% confidence interval, 0.801–0.950) in predicting severe posthepatectomy liver failure and achieved well-fitted calibration ability. The predictive performance of the artificial neural network model for severe posthepatectomy liver failure outperformed the traditional logistic regression model and commonly used scoring systems. Moreover, stratification into 3 risk groups highlighted significant differences between the incidences and grades of posthepatectomy liver failure.

Conclusion

The artificial neural network model accurately predicted the risk of severe posthepatectomy liver failure in patients with hepatocellular carcinoma who underwent hemihepatectomy. Our artificial neural network model might help surgeons identify intermediate and high-risk patients to facilitate earlier interventions.

Introduction

Globally, hepatocellular carcinoma (HCC) is the sixth most common form of cancer and the fourth leading cause of cancer related death.¹ Major hepatectomy is the main type of radical therapy used for the treatment of HCC patients with extensive invasion into liver tissues.2, 3, 4 Although advances in surgical techniques for liver resection and perioperative care have greatly improved prognoses and treatment outcomes, posthepatectomy liver failure (PHLF) is a worrisome complication after major hepatectomy and is the leading cause of postoperative mortality.5, 6, 7, 8, 9 In addition, PHLF is the major cause of prolonged hospital stays, increased costs to patients and hospitals, and decreased long-term prognosis for HCC patients who have undergone hepatcetomy.⁷ Currently, techniques to help reduce the morbidity and mortality of PHLF are urgently needed by concerned hepatobiliary surgeons. One essential step to facilitate these needs and avoid PHLF and postoperative mortality is to increase the accuracy of preoperative assessments of liver functional reserve, particularly for HCC patients with potential impairment of hepatic function.

Generally, the development of PHLF is mainly determined by both the quality and quantity of hepatocytes after hepatectomy.¹⁰ Currently, there are many commonly used scoring systems for the determination of the quality of the liver, including (1) Child-Pugh grade,^11,12 (2) model for end-stage liver disease (MELD),¹³ (3) albumin-bilirubin (ALBI),^14,15 (4) platelet-albumin-bilirubin (PALBI),^16,17 (5) fibrosis index based on the 4 factor (FIB-4),^18,19 and (6) the aspartate aminotransferase to platelet ratio index (APRI) scoring system.20, 21, 22 However, these scoring systems are based on clinical symptoms and blood tests and do not take into consideration a number of certain important components of the liver and associated tissue such that many limitations have been revealed.11, 12, 13, 14, 15, 16, 17, 18, 19, 20, 21, 22 For example, the Child-Pugh grade scoring system includes subjective criteria such as for ascites and hepatic encephalopathy, ultimately limiting its ability to predict early complications from posthepatectomy.^11,12 The MELD scoring system was designed for patients with advanced cirrhosis and is inappropriate for most HCC patients.¹³ The ALBI and PALBI scoring systems failed to eliminate the “ceiling effect” of hyperbilirubinemia and hypoproteinemia on liver function assessment.14, 15, 16, 17 The FIB-4 and APRI scoring systems have mainly been used as noninvasive diagnostic tools for the assessment of fibrosis and cirrhosis; however, they require additional study to fully elucidate their impacts on liver quality.18, 19, 20, 21, 22 Moreover, indocyanine green retention test at 15 minutes (ICG-R15) is another tool that has been feasibly used to evaluate liver quality, but it is mainly used in Japan and is not as well accepted in other countries.^23,24 Therefore, these liver function assessment methods, when used alone, may not be reliable for the accurate prediction of PHLF.

Based around recent efforts, many studies have confirmed that computed tomography (CT) volumetric imaging could be used to effectively assess liver functional reserve and that this may be especially useful for patients with HCC who underwent major hepatectomy.^25,26 Future liver remnant (FLR) volume has also been used to calculate using 3 dimensional CT reconstruction technology, and may be effectively indicate the quantity of remnant liver after resection^25,26; whereby, the lower the FLR value the higher the risk of serious complications after surgery.^27,28 Nevertheless, FLR volume measurements did not exactly assess the quality of the hepatic parenchyma; therefore, this measure is not appropriate for assessing PHLF risk in HCC patients with damaged livers.^29,30 Therefore, the FLR volume measurements should be combined with methods that facilitate evaluations of liver function to improve ability to predict severe PHLF.

Owing to the lack of a specific and practical model for predicting the risk of severe PHLF, development of a clinically applied model that combines both the quality and quantity of the FLR becomes desirable. Among all presently available methods, artificial neural network (ANN) models are mathematical tools that effectively imitate the structure of biological neural networks.³¹ ANN models possess the features including parallel-distributed data storage and data processing, good fault-tolerance, the ability to deal with high levels of non-linearity, are strongly self-adaptive, are self-organizing, and have self-learning abilities.³¹ Of the many types of assessments of disease risk, ANN models have proven to be more effective than traditional discriminant analysis and as such have become an often used alternative model and in some cases become a new standard.^32,33 Nevertheless, thus far little research has examined the efficacy of use of ANN models for patients afflicted with HCC. Therefore, we sought to develop and assess the accuracy and reliability of a novel ANN model that would combine information for the quality and quantity of hepatocytes to predict preoperative individualized risk of severe PHLF for patients with HCC who underwent hemihepatectomy. Further, we sought to compare the discriminative ability of the ANN model with results from traditionally oriented logistic regression (LR) and commonly used scoring systems.