In this study, we have summarized the natural disease progression and survivals of patients with HCC-derived spinal metastases and explored for potential prognostic factors for prediction of survival to derive a novel individualized prediction tool.
The incidence of HCC in Southeast Asian countries, including Thailand, is high [11]. Hepatitis B and C virus infection and alcoholic cirrhosis were important etiologies in our patients, which differed from a 36 patients case series from China, which were all derived from hepatitis B virus infection [5]. From our 10-year study, only six percent of patients with HCC were found to have spinal metastases with a median survival time of 79 days, which was relatively short compared with that of other common metastatic tumor to the spine such as lung cancer (11.3 months) [12], breast cancer (21.7 months) [13], prostate cancer (58.3 months for hormone naïve and 5 months for hormone refractory [14, 15]), thyroid cancer (15.4 months) [16], and cholangiocarcinoma (3 months) [17]. In comparison with other reports on the survival of patients with HCC-derived spinal metastasis [6, 18, 19], the median survival time after metastasis was also shorter in our study. This could be explained by the lower rate of primary surgical resection and higher rate of palliative treatment and best supportive care in our study. As the median survival time after the diagnosis of spinal metastases was much shorter than the median time from diagnosis of primary HCC to the diagnosis of spinal metastases, the existence of undetectable occult spinal metastases was plausible.
Several studies have reported a variety of prognostic factors for survival in patients with HCC with spinal metastases. Interestingly, most of these factors were either patient-related (i.e., Eastern Cooperative Oncology Group or ECOG and KPS) [6, 20–22], liver-related (i.e., serum albumin level, serum lactate dehydrogenase or LDH, and Child-Pugh classification) [6, 21, 23, 24], or metastatic-related (visceral metastasis, other extrahepatic metastasis other than bone metastases) characteristics [22–25]. For tumor-related or intervention-related factors, primary HCC control or response to HCC treatment (i.e., response to radiotherapy, previous resection of primary HCC) were reported to be associated with patient survival after the diagnosis of spinal metastases [6, 20, 22, 24]. Previous scoring systems (i.e., Tomita score and Revised Tokuhashi score) have also been reported to be capable of survival prediction in this domain of patients [5, 23, 26]. However, other potentially important HCC-related factors such as the number of tumors (or multifocality) and tumor size have never been investigated as prognostic factors in patients with spinal metastasis. In this study, we included the number of tumors as a candidate predictor, as it was significantly associated with both recurrence and metastasis in HCC, even after adequate primary HCC control [27]. In contrast, the tumor size was not modeled as only a few patients (14.5%) had tumor size less than 5 cm, and the survival distribution between groups with tumor size smaller or larger than 5 cm was not significantly different. To allow for comprehensive and accurate survival prediction, our model incorporated all relevant aspects of the cancer, which were patient-, liver-, and tumor-related factors.
For patient-related factors, age and KPS were identified as significant prognostic factors in our study and were included as predictors in the model. Advanced age has been widely reported to be associated with poor survival outcomes in HCC, regardless of the staging or therapeutic modes [28, 29], but has never been explored in patients with HCC with spinal metastases. Previous studies used a variety of age cut points, ranging from 60 to 70 years old [27–30]. Our study is the first one to demonstrate the effect of old age (> 60 years) on the survival of patients with HCC-derived spinal metastases. The patient’s performance status, either via ECOG or KPS, was consistently identified as an influential survival factor in HCC patients with spinal metastases [6, 23] and were included as predictors in two recently developed prediction models: HCC-SM GPA by Rim et al. [24] and another scoring system by Uei and Tokuhashi [6].
For liver-related factors, only serum total bilirubin was found to be an independent survival predictor in our patient cohort. Serum bilirubin had never been independently explored as a prognostic factor in HCC-derived spinal metastasis, as all previous studies only examined the effect of serum albumin level or Child-Pugh classification as a whole [6, 21, 23, 24]. In our analysis, we separated the component of Child-Pugh classification to examine their independent effect on patient survival. Both serum albumin and serum bilirubin were found to be significant predictors in the multivariable model, but only serum bilirubin remained in the reduced model, which could be explained by the limited study size and that the effect estimates of serum bilirubin were much larger. Commonly, hyperbilirubinemia is a dominant marker of liver damage or liver failure[31]. It was recently found that elevated serum bilirubin (≥ 1.5 mg/dL) was associated with HCC aggressiveness, increased risk of portal vein thrombosis (PVT), and lower survival, regardless of the tumor size. Moreover, patients with hyperbilirubinemia were found to have lower platelet counts, lower serum albumin, higher aspartate aminotransferase (AST), and higher alkaline phosphatase (ALP) levels than patients with normal bilirubin levels [32]. In our data, both serum albumin level and PVT revealed significant trends across the ordered groups of total bilirubin (P-value from non-parametric test for trend 0.021 and 0.035, respectively).
For tumor-related factors, the multifocality of HCC was revealed to be another independent prognostic factor for survival in HCC-derived spinal metastasis. Multiple primary tumors reflect intrahepatic metastasis. Even when visible tumors are adequately resected, the remaining small metastases may still lead to recurrence and metastasis of HCC [33]. Metastatic-related factors such as visceral organ metastasis or metastases to the major internal organ, number of extraspinal bone metastases, and number of vertebral columns involved in spinal metastasis were not identified as significant predictors in our study, which was in concordance with other studies [6, 7]. Due to the limitations in terms of study sample size and incomplete data availability, we did not include intervention-related factors such as primary HCC control modality, treatment with bone modifying agent, and sorafenib as predictors. In addition, more than 2/3 of the patients in this study (68.1%) did not receive definite HCC treatment and only three patients (4.4%) were offered surgical resection, and most patients received best supportive care or palliative treatment. It was quite evident from our exploratory analysis that prognostic factors for HCC survival at diagnosis of the cancer still affected the patient’s life expectancy after spinal metastases occurred.
Most clinical decision tools for survival prediction in patients with metastatic spinal tumors were developed using the conventional Cox’s proportional hazard regression model because of its statistical simplicity and comprehensible concept [6, 8, 9, 34]. However, as a semi-parametric method, the baseline hazard function cannot be directly estimated from the model itself without conditioning the estimated regression coefficients, and the rigid proportional hazard assumption must be fulfilled for valid estimates. In addition, the model was specifically designed to assess the effect of each prognostic covariate on the change in the patient’s hazard function; it was not intended to predict the survival function of each patient [35]. For derivation of Cox’s prognostic model, weighing of coefficients was generally performed to generate a score for each patient from a combination of predictors. The score was subsequently categorized into different risk groups at arbitrary cutoff points for clinical application. Thus, precise individual prediction could not be performed using the Cox’s model [36].
In this study, we employed an alternative approach to directly estimate both the baseline hazard function and the individual survival function via the RP flexible parametric survival model. The RP model has been proven to yield more accurate predictions and more precise calibration in survival prediction [36, 37]. In deciding the appropriate treatment plan for HCC patients with spinal metastasis, an accurate survival prediction at clinically-relevant time points is needed for each specific patient, especially for surgical management (either excisional or palliative). Applying predictions based on categorized risk groups might be considered unsophisticated as the splitting of an initially wide range of continuous prediction results in significant loss of information and certainly impairs the accuracy of overall prognostication [36]. With the HCC-SM CMU model, we proposed a novel practical tool for individualized prediction of survival probabilities in HCC patients with spinal metastases. The HCC-SM CMU model incorporates four significant clinical predictors to approximate survival probabilities and their confidence intervals for a specific patient at multiple time points.
In comparison to the widely used Tomita and Revised Tokuhashi score, our model carried higher discriminative ability in terms of C-statistics, as the other scores were not derived from the full cohort of patients with HCC-derived spinal metastasis. With four simple and readily available predictors, the application of the HCC-SM CMU model in practice would result in lower false positive cases than both the Tomita and Revised Tokuhashi scores. As the model was intended to be used for guiding the need for major surgical operations, a higher specificity was indeed more important than the sensitivity. However, all differences in diagnostic indices were not statistically evident because of the limited statistical power. The most outstanding point of the HCC-SM CMU model over the traditional scores, including the HCC-SM GPA and the newly derived prognostic scoring system by Uei and Tokuhashi was that it allowed for a wider range of predictions as seen from the score chart. For each clinically-relevant time point, the HCC-SM CMU model exhibited 36 predicted survival probability values according to the individual characteristic pattern. These individual predictions could play an important role in risk communication and decision making from both the patient and physician perspectives [38].
Our study has several limitations. First, the model was derived from a small sized patient cohort with retrospective data collection. Even though the number of censored observations was minimal, it is questionable whether the available number of events and total follow-up duration would be sufficient for model derivation with a flexible parametric model, which requires more parameters for the natural cubic spline function. Thus, to prevent model overfitting and optimism, we limited the number of final predictors and used a bootstrap resampling procedure to assess the presence of optimism and generate the shrinkage factor for further validation [39]. In terms of missing data, both the multiple imputation method and complete-case analysis were used to estimate the multivariable prediction model. Second, all predictors were modeled as categorical variables instead of continuous variables. This could result in the loss of information and an overoptimistic model in case data-driven approaches have been used. In our study, all predictors were categorized according to prior clinical scoring systems and generally accepted cut points to prevent optimistic model performance [40, 41]. Third, the model calibration was poor in specific ranges of predicted probabilities, and the discriminative ability was only fair to acceptable. Regarding calibration, careful interpretation of predicted survival probabilities was suggested, especially in the second risk quantile, where the survival probabilities were overestimated. In contrast, patients with predicted probabilities above 60% at 180 days would have an even higher chance of surviving beyond 6 months and therefore be a good candidate for surgical management. Prior to the clinical implication of the HCC-SM CMU model, a prospective external validation study with a larger HCC population is warranted.