Uveal melanoma (UM) is the most common intraocular primary malignancy in adults involving the choroid, iris and ciliary body, and has a high tendency to metastasize, resulting in high mortality (1–3). Approximately half of the patients develop metastatic disease during a long follow-up (1, 2). Currently, there is no proven effective treatment for patients with metastasis. Once metastasis is detected, the patient survival rate rapidly decreases to approximately 15% at one year, and the median survival time is estimated to be 4 to 15 months (4). Some clinical trials are being conducted with patients with metastasis, which may provide some benefits for patients (5, 6). Therefore, early screening of high-risk patients and early detection of metastasis are very important. Moreover, establishing risk stratification and making different follow-up plans accordingly can also save many medical resources, reduce unnecessary invasive examinations and testing expenses, and thus result in rational allocation of medical resources.

Worse prognostic indicators for UM include 1) clinical factors, such as older age at presentation, male sex, larger basal tumor diameter, high tumor thickness, ciliary body involvement, and extraocular spread (7, 8); 2) histopathological factors, such as epithelioid melanoma cytomorphology, certain extravascular matrix patterns such as closed loops (9), a high mitotic count, and high tumor microvascular density (8); and 3) genetic abnormalities, including chromosomal changes (chromosome 3 loss, 8q gain, 1p loss and 6p gain) (8), class 2 gene expression profiles (10), and other biomarkers (BAP1, SF3B1, EIF1AX, and PRAME) (11, 12). Most of the prognostic prediction models have been established based on these risk indicators, such as the TNM staging system, the gene expression profile (10, 13), the Liverpool Uveal Melanoma Prognosticator Online (14), the Cancer Genome Atlas Classification (15) and other models (16–19). Most of the present prognostic models are relatively complicated and less feasible. The high expenses of these tests may also be limitations for some patients, especially those living in developing regions and countries.

To our knowledge, no prognostic models have been analyzed and established based on Asian patients. Asian patients, compared to the Caucasian population, have some unique features, such as a lower incidence probability, earlier onset age, and larger tumor diameter (20, 21). The two largest clinical prognosis studies in China showed that Chinese patients have a younger onset age and no sex difference and tend to have a better prognosis (22, 23). Studies from other Asian countries (Japan (24), India (25), Korea (26) and Singapore (27) also showed similar results.

Based on the long-term follow-up results of the patients in our hospital, we established a novel and simple nomogram prediction model to predict the survival of UM patients. Two public databases from Western countries were analyzed for external verification.

The TNM staging system is used worldwide for UM and integrates the basal tumor diameter, tumor height, tumor location, extraocular extension, lymph glands and distant metastasis and could thus be used as a prediction model. The Cancer Genome Atlas Class groups tumors as A, B, C or D based on their chromosome status and can also be used to predict metastasis and death (15, 33). The class 1 and class 2 gene expression profiles are genetic assessment tools that are widely applied in Western countries, especially in the United States. Based on genetic analysis, the gene expression profile can distinguish patient prognosis with a significant difference in that nearly all metastatic deaths fall into the class 2 category, and it has been validated on multiple independent datasets (13, 34). Another promising estimating model is the Liverpool Uveal Melanoma Prognosticator Online, which combines pathological, clinical and genetic data to provide individual patients with an estimated survival time as well as the risk of metastasis (14). This model has been validated with several international cohorts and can serve as a valuable tool for predicting all-cause mortality (35, 36). In addition, there are several other prognostic models, such as the artificial neural network by Kaiserman (16), the nomogram-SEER model by Zeng (17), the parsimonious model by Damato (18), and the Markov multistate model by Eleuteri (19). The main characteristics of these models are listed in Table 4. Although gene and chromosome testing increase the accuracy of predictions, the high expense and intricate testing technology may limit their wide application in developing regions and countries. Moreover, from Table 4, we found that these models were developed mostly with the Caucasian population.

Our prediction models screened out only five clinicopathologic variables and did not include gene testing or chromosome typing. However, this model still achieves rather high prediction accuracy and showed even better discrimination and calibration than those of the TNM staging system in our patients. This result suggested that these clinicopathologic features, even without the genetic phenotypes, highly represent the inherent characteristics of the disease, further indicating that the gene phenotypes may have been outwardly expressed by these clinicopathologic features in our patients. Thus, the prediction accuracy was not sacrificed when evaluating without genetic testing. Our model was very easy to operate and has low requirements for testing equipment and a low testing cost; hence, it is highly feasible for use by patients from less medically developed countries and regions. Moreover, we conducted validation in the TCGA and SEER databases, and the model showed good discrimination and calibration in these two databases, which indicated that our nomogram model is also suitable for Western patients.

Comparing the baseline of our patients with the Western databases, the clinicopathology and prognosis of our patients were different: our patients had a younger onset age, less male tendencies, less non-spindle cell type, less ciliary body involvement and less extra-scleral extension, and our patients showed a better survival trend. It is inferred that the better prognosis may be related to the inherent genetic features. A study sequenced the whole genome of Chinese patients and found that the GNA11 mutation rate was only 21.5% in those patients, which was much lower than the reported rate of 32.6% in a Caucasian population; instead of BAP1, SF3B1, and EIF1AX, their patients with HIF1A and FOXO1 mutations exhibited worse OS (37). This finding suggests that there may be inherent genetic differences between the two groups of patients.

Based on this model, the risk stratification displayed an obvious distinction between the low- and high-risk groups in both the training and validation cohorts. This result suggests that the model could provide an excellent differential ability in prognosis prediction. Thus, high-risk patients could be easily screened out and closely followed up to detect and treat metastasis earlier. Furthermore, we found that the curves in the training cohort were very flat and smooth, especially the one indicating the low-risk group, while the corresponding curves in the TCGA and SEER cohorts showed a decreasing tendency. This result indicated that this model reflected the prognosis of Chinese patients well, while there may be other independent variables affecting the prognosis of Caucasian patients. Therefore, it may also serve as proof that there are physiological differences between Asian and Caucasian patients.

There are some limitations of this study: 1. Although we used two external databases, the follow-up time of these two databases was limited, so we can only verify the 5-year survival rate instead of the long-term prediction effect of the 10-year survival rate. 2. We only selected patients with complete information, especially in the SEER database, which could have potentially introduced selection bias. 3. Due to the characteristics of the retrospective study, we didn’t compare the efficiency of our model with other UM prediction models. In the future, some prospective studies will be carried out to further verify the efficacy of the present model, and its comparisons with other UM prediction models will be designed as well.

Model II has the following strengths: 1. This model was established based on Asian patients with UM, and also proved applicable to the Caucasian population, thus indicating its wide scope of application. 2. The model requires only a few variables, which are easy to obtain clinically. It could not only provide the estimated probability of survival for each patient, but also differentiate the low- and high-risk patients (with <170 or ≥170 cores) and thus make different follow-up plans accordingly. For example, we could reduce unnecessary invasive examinations and testing expenses for low-risk patients, and give close follow-ups or preventive treatments, such as clinical drug trials, to patients with high risk. In this way, the model is simple to operate and will contribute to rational allocation of medical resources.

In conclusion, we developed a simple and easy-to-promote nomogram model to predict the OS of patients with UM based on our Chinese population, and it was more accurate than the TNM staging system. Our prediction model also showed high discrimination and good calibration when verified by two Caucasian databases, suggesting that the model is suitable for a wide population.

The raw data supporting the conclusions of this article will be made available by the authors, without undue reservation.

The studies involving human participants were reviewed and approved by Ethics Committee of Eye & ENT Hospital of Fudan University. The patients/participants provided their written informed consent to participate in this study.

Conception and design (HY and JQ); Data collection (HY, BX, YB, KX, JGu., RZ, HR, YY, and JQ); Data analysis and interpretation (JGa.); Manuscript preparation (HY, BX, and JQ); Obtained funding (KX). All authors contributed to the article and approved the submitted version.

Project supported by the Shanghai Committee of Science and Technology, China (Grand No. 20Y11911200). The funding had no role in the design or conduct of this research.

The authors declare that the research was conducted in the absence of any commercial or financial relationships that could be construed as a potential conflict of interest.

All claims expressed in this article are solely those of the authors and do not necessarily represent those of their affiliated organizations, or those of the publisher, the editors and the reviewers. Any product that may be evaluated in this article, or claim that may be made by its manufacturer, is not guaranteed or endorsed by the publisher.