A prognostic nomogram in AFP-negative hepatocellular carcinoma based on LASSO Cox regression

Purpose Hepatocellular carcinoma (HCC) patients with alpha-fetoprotein (AFP)-negative (<8.78 ng/mL) have special clinicopathologic characteristics and prognosis. The aim of this study was to apply a new method to establish and validate a new model for predicting the prognosis of HCC patients with AFP-negative. A total of 410 AFP-negative patients with clinical diagnosed with HCC as a primary cohort; 148 AFP-negative HCC patients as an independent validation cohort. In primary cohort, independent factors for overall survival (OS) by LASSO Cox regression were all contained into the nomogram1; by univariate and multivariate Cox hazard analysis were all contained into the nomogram2. Nomograms performance and discriminative power were assessed with concordance index (C-index) values, area under curve (AUC), Calibration curve and decision curve analyses (DCA). The results were validated in the validation cohort. nomogram1 was clinically useful. Moreover, patients were divided into three distinct risk groups for OS by the nomogram1: low-risk group, middle-risk group and high-risk group, respectively. Conclusions Novel nomogram based on LASSO Cox regression presents more accurate and useful prognostic prediction for patients with AFP-negative HCC. This model could help AFP-negative HCC facilitate a personalized prognostic evaluation.


Introduction
HCC is the fifth most common malignancy worldwide and the third most associated with all cancer deaths [1]. As with other cancers, the prognosis for early HCC is better than late stage [2]. Since the identification of alpha-fetoprotein (AFP) in 1970s, it has been the only serologic marker that is widely used for the HCC diagnosis [3]. However, the diagnostic power of AFP has been continuously questioned and debated. For example, elevated serum AFP was only observed in 60-70% of overall HCC patients, while the proportion was merely 33-65% regarding patients harboring HCCs of <3 cm in diameter [4,5]. HCC patients with negative serum AFP have special clinicopathologic characteristics and prognosis, they have higher tumor differentiation, earlier TNM staging, smaller tumor size, and higher survival rates [6].
Many staging systems have been employed to predict how HCC patients will respond over time, such as the BCLC, ALBI, Child-Pugh and TNM staging system. However, studies have shown that the traditional staging system is flawed to varying degrees. The BCLC staging system was reported to have the greatest potential in predicting HCC in AFP negative patients [7]. However, evidence has shown that the classification of the BCLC score is limited to the advanced stages of HCC [8]. Child-Pugh does not consider tumor-related factors, which are important for the prognosis of HCC patients [9], as the prognosis of HCC patients, tumor-related factors are crucial. ALBI model incorporates few factors, clinical indicators are easily available and easy to apply, but there are no tumor-related indicators to evaluate [10], as a model to evaluate the prognosis of HCC needs to be validated in a multicenter large sample. The TNM staging system is easy to use and is considered to be the best staging system for solid tumors, but its effect on the staging and prognosis of HCC is debatable because it only considers tumor characteristics but not liver function, which usually plays an important role in the prognosis of HCC patients [11]. Furthermore, these systems are not specifically designed to predict outcomes of HCC when patients are AFP-negative. Recently, nomograms for prediction of survival and recurrence of HCC in AFP-negative patients after radical resection have been developed in two previous studies [12,13], respectively. However, HCC patients with transarterial chemoembolization (TACE) and radiofrequency ablation (RFA) were not included in both studies. At the same time, most of the current models are limited to Cox univariate and multivariate analysis for risk factor screening, which is unconducive to the small sample but multi-indicators models screening. Therefore, the aim of our study was applying a new method to establish and validate a new model that combines clinical pathological factors, biochemical indicators, for predicting the prognosis of AFP-negative HCC. In addition, a comparison between the constructed nomograms and traditional staging systems was conducted to determine whether the nomograms provided more accurate prediction in prognosis.

1.Patient selection
We performed a retrospective cohort study on 458 AFP-negative (<8.78 ng/mL) HCC patients using data from the Beijing Ditan Hospital between January 2008 and December 2016. Diagnostic criteria of HCC were consistent with previous studies [14]. The HCC diagnosis data included biopsy, radiology. First, we selected patients based on hepatic angiography, pathology in combination with ultrasonography, computed tomography (CT), and magnetic resonance imaging (MRI). Next, we only

2.Laboratory Measurements
Patients are routinely examined at the first visit. Data provided include: gender, When laboratory values at the time of HCC diagnosis were higher than clinical normal, it was classified as elevated. The tumor was staged using the BCLC staging system and liver function was scored using Child-Pugh.

3.Statistical analysis
Statistical analysis was done using SPSS 24 In this study, we used the primary cohort to plot nomogram1 based on LASSO Cox regression [15], and used the primary cohort to plot nomogram2 based on multivariate COX analysis. Based on the established nomogram1 and nomogram2, the C-index and calibration curves were derived based on COX regression analysis, NRI and IDI scores were performed, and decision curves were plotted. Nomogram1 was found to be superior to nomogram2 in the primary and validation cohorts. The total score for each patient was calculated based on nomogram 1, and three groups of patients with high, medium, and low prognostic risk (based on the total score) were divided according to interquartile values. The Kaplan-Meier curve was applied in MedCalc software, and the risk group was compared using the three-point factor, log-rank test, and the two-tailed P value < 0.05 was statistically significant.

1.Basic characteristics
In the primary cohort, 410 patients with AFP-negative HCC met the inclusion and exclusion criteria and were included in this study. A total of 148 HCC patients with negative AFP were included in the validation cohort. The clinical characteristics of patients in the two independent cohorts are shown in Table 1 (Table 1). In the primary cohort, 224 patients died from cancer within 5 years.  Continuity cut point 0.05NRI (-0.037), subtype cut point 0.5NRI (-0.037), IDI: 0.006 (p=0.29). We get similar results in the validation cohort. All suggest that nomogram1 is better than nomogram2. The decision curve analysis of nomogram 1 and nomogram 2 is also meaningful ( Figure 5). In the primary cohort the cue differences were small, but in the validation cohort nomogram1 and nomogram2 were better at predicting OS than either the allpatient death scenario or the no-patient death scenario if the patient threshold probability was > 20%. Moreover, the net benefit is comparable; in this range, the predicted OS of nomogram 1 is more advantageous than nomogram 2.

5.Application of the Nomogram model for risk stratification
Based on the nomogram1 we developed in this study, we subdivided the patients into low-risk, middle-risk, and high-risk groups, and the AFP-negative HCC patients showed good prognostic classification in both the primary cohort and the validation cohort. In the primary cohort, there were 122 cases in the low-risk group, 175 cases in the middle-risk group, and 113 cases in the high-risk group. Intergroup

Discussion
It is important to have an accurate prognosis of the tumor after relevant treatment.
In this study, we identified 3 years and 5 years of independent risk factors for OS in AFP-negative HCC patients in the primary cohort, and established novel, effective, and verified nomograms1 to predict individuals at 3 years and 5 years OS.PVTT, ascites, HGB, γ-GGT and CRP have been included in the nomogram1. In addition, nomogram1 shows a higher discriminatory power, and the C-index of the two independent cohorts are 0.708 and 0.752, respectively. In the subsequent study, the calibration curve and decision curve analysis were used to assess the predicted precision and clinical utility of the nomogram1. Our model showed a better net benefit and higher consistency between the nomogram prediction and the actual observation.
Previous studies have demonstrated the prominent role of tumor burden and grade, liver function, degree of hepatic dysfunction, and performance status in the prognosis of HCC. For many years, the traditional Child-Pugh rating system is the most widely used method for assessing liver function and predicting therapeutic efficacy [16]. Our study also found that the C-index (0.629) and AUC (0.672) of Child-Pugh in the primary cohort were significantly higher than ALBI, TNM, and BCLC. We also obtained the same results in the validation cohort. However, AFP-negative HCC patients have special clinicopathologic characteristics and prognosis, they have higher tumor differentiation, earlier TNM staging, smaller tumor size, and higher survival rates [6]. Most models are built without taking into account the specificity of liver biology function in AFP-negative HCC and do not fully reflect an accurate prognosis.
The above studies indicate that our model is more accurate for the prognosis of AFPnegative HCC patients. We found C-index (0.708), AUC (0.736) and time-dependent ROC curves of our model was better than the other models, such as Child-Pugh, BCLC, ALBI, TNM. The calibration curve for probability of OS showed good homogeneity between prediction by our model and actual observation. DCA demonstrated that our model was clinically useful. We get similar results in the validation cohort.
A variety of nomograms have been developed to predict the prognosis of certain cancers and have been shown to be more accurate than traditional staging systems.
However, most of them are limited to COX univariate and multivariate risk factor screening, which is not conducive to small sample size, multi-indicator model screening [17]. LASSO Cox regression analysis constructs a penalty function to obtain a more refined model. It compresses the regression coefficients (the sum of the absolute values of the mandatory coefficients is less than a fixed value) and sets some regression coefficients to 0. LASSO Cox regression analysis can not only solve the problem of over-fitting, but also extract useful features effectively [15]. Meanwhile, LASSO Cox regression is more applicable in decisions with more clinical indicators.
In our study, we find that the model constructed by applying the factors filtered by LASSO Cox regression has better accuracy and resolution than the model constructed by traditional Cox regression. In the primary cohort, we find the C-index (0.708) and AUC (0.736) of nomogram1 is superior to the C-index (0.706) and AUC (0.714) of nomogram2. We got better results in the validation cohort.  (Figure7). CRP is synthesized in the liver and is secreted into the plasma as a pentamer, belonging to the family of pentraxins, together with serum amyloid protein [18]. CRP can predict overall survival and recurrence rates after hepatectomy in patients with HCC patients [19]. Meanwhile, CRP was independently associated with OS in non-surgical HCC patients [20].The ascites, jaundice, hepatic encephalopathy, and liver failure may induce by the formation of PVTT and thus might account for, at least in part, the difference in prognostic factors between HCC with and without PVTT [21].Previous studies indicate that increased γ-GGT is closely associated with the development of HCC because of its positive correlation with tumor size and poor OS [22].Anemia is common in cancer patients. HGB levels have been shown to have an impact on survival both before and during anti-cancer therapy [23]. The pre-treatment anemia in HCC patients was found to be 7.0%, which is less than the 12.8% of cancer-related anemia [24]. The causes of anemia in HCC patients include nutritional deficiencies, hemolysis, blood loss, and tumor cell infiltration of the bone marrow [25]. In addition, chronic liver injury can also lead to anemia in HCC patients [26]. It has also been shown that downregulation of iron-regulated genes, including heparin, ceruloplasmin, transferrin, and transferrin receptors, disrupts the systemic iron homeostasis, leading to anemia in patients with HCC [27]. It has been demonstrated that HGB is an independent prognostic factor in patients with HCC, which is similar to the results we obtained [24]. Ascites is the hallmark of portal hypertension. Besides, ascites may also be associated with large tumor burden and vascular invasion of HCC, suggesting that the cause of ascites could be attributed to both tumoral and cirrhotic factors [28]. Studies have shown that ascites reduces longterm survival in patients with liver cancer [29]. It is similar to the results of our current study. Tumor size is one of the most important parameters of tumor burden.
However, evaluation of tumor number failed to predict the prognosis of HCC patients in the present study, although other studies have demonstrated its prominent role in prognostic prediction [30]. Meanwhile, tumor markers have not been included in this study. After research and repeated comparison, our nomogram exhibited superior discrimination ability for the prediction of prognostic in AFP-negative HCC patients.
Hence, our novel nomogram could be used to guide routine follow-up for patients.
PVTT, ascites, HGB, γ-GGT and CRP should be given great importance in AFPnegative HCC patients. In addition, patients given a high score by the nomogram should undergo more high-end imaging examinations, such as MRI or CT exams, and the interval time of follow up should be reduced, even if the last test results have no causes for concern.
Although our nomogram performed well, several limitations need to be addressed.
Firstly, the nomogram was derived from data collected by one agency and not validated internally due to the difficulty of obtaining such data; in the validation cohort the follow-up time was shorter, and close monitoring and five-year follow-up data are still required for patients in the validation cohort. Secondly, because the present study was a retrospective study for predicting anticipated future performance, our results need to be confirmed by prospective cohort studies. Thirdly, since only three of the patients we included underwent resection, is our nomogram applicable to patients who underwent radical resection of AFP-negative patients remains uncertain.
In conclusion, we applied new methods to develop and validate nomogram to predict the OS in AFP-negative HCC patients. Nomogram1 presented in this study is statistically easier than previous model screening methods and more accurate than ALBI, TNM, Child-Pugh, BCLC, and provides a useful tool for prognosis. In order to promote the use of nomogram1 in other groups, additional validation of data from other institutions is needed.