Pan-cancer analysis of prognostic metastatic phenotypes

Abstract

Although cancer is highly heterogeneous, all metastatic cancer is considered American Joint Committee on Cancer (AJCC) Stage IV disease. The purpose of this project was to redefine staging of metastatic cancer. Internal validation of nationally representative patient data from the National Cancer Database (n = 461 357; 2010-2013), and external validation using the Surveillance, Epidemiology and End Results database (n = 106 595; 2014-2015) were assessed using the concordance index for evaluation of survival prediction. A Cox proportional hazards model was used for overall survival by considering identified phenotypes (latent classes) and other confounding variables. Latent class analysis was performed for phenotype identification, where Bayesian information criterion (BIC) and sample-size-adjusted BIC were used to select the optimal number of distinct clusters. Kappa coefficients assessed external cluster validation. Latent class analysis identified five metastatic phenotypes with differences in overall survival (P < .0001): (Stage IVA) nearly exclusive bone-only metastases (n = 59 049, 12.8%; median survival 12.7 months; common in lung, breast and prostate cancers); (IVB) predominant lung metastases (n = 62 491, 13.5%; 11.4 months; common in breast, stomach, kidney, ovary, uterus, thyroid, cervix and soft tissue cancers); (IVC) predominant liver/lung metastases (n = 130 014, 28.2%; 7.0 months; common in colorectum, pancreatic, lung, esophagus and stomach cancers); (IVD) bone/liver/lung metastases predominant over brain (n = 61 004, 13.2%; 5.9 months; common in lung and breast cancers); and (IVE) brain/lung metastases predominant over bone/liver (n = 148 799, 32.3%; 5.7 months; lung cancer and melanoma). Long-term survivors were identified, particularly in Stages IVA-B. A pan-cancer nomogram model to predict survival (STARS: site, tumor, age, race, sex) was created, validated and provides 13% better prognostication than AJCC: 1-month concordance index of 0.67 (95% confidence interval [CI]: 0.66-0.67) vs 0.61 (95% CI: 0.60-0.61). STARS is simple, uses easily accessible variables, better prognosticates survival outcomes and provides a platform to develop novel metastasis-directed clinical trials.

Keywords: cancer; death; metastasis; phenotype; prediction.

Figure 1.. Distributions of metastatic cancer

Figure 1A shows relative percent of metastatic distributions to the 4 main sites among different cancers analyzed, in a heat map, with red being frequent and blue/purple being infrequent. Certain cancers had propensity to spread only to one site (e.g. pancreas to the liver); in contrast, others had propensity for heterogeneous metastatic patterns (e.g. melanoma). Figure 1B-1F. Latent class analysis identified five principal metastatic phenotypes. Heat maps, with red being frequent and blue/purple being infrequent, display frequency of cancer patients among each phenotype. Latent class analysis identified five phenotypes: (IVA) nearly-exclusive bone-only metastases, commonly seen in lung, breast, and prostate cancer (n = 59,049, 12.8%); (IVB) predominant lung metastases, commonly seen in cancers of the breast, stomach, kidney, ovary, uterus, thyroid, cervix, soft tissue (n = 62,491, 13.5%); (IVC) predominant liver/lung metastases, commonly seen in cancers of the colorectum and pancreas (which create the majority of this phenotype), as well as cancers of the lung, gallbladder, esophagus, and stomach (n = 130,014, 28.2%); (IVD) bone/liver/lung metastases predominant over brain, commonly seen in lung and breast cancer patients (n = 61,004, 13.2%); and (IVE) brain/lung metastases predominant over bone/liver, commonly seen in lung cancer and melanoma (n = 148,799, 32.3%). Certain cancers almost always followed one phenotypic pattern, e.g. colorectal and pancreatic cancer in phenotype IVC, and prostate in phenotype IVA. In contrast, other cancers could present with any of these five phenotypes, most commonly kidney and esophageal cancer.

Figure 2. STARS model to predict phenotype and survival

Figure 2A. Kaplan-Meier survival plots of the metastatic phenotypes. Median survival time for phenotypes IVA-IVE were as follows: 12.6 months, 11.4 months, 7.0 months, 5.9 months, and 5.7 months. Phenotypes IVA and IVB had similar survival, as did phenotypes IVC, IVD, and IVE. Figure 2B. Nomogram to predict survival among metastatic cancer patients. The STARS nomogram includes site, tumor, age, race, sex to predict 3-year and 5-year survival. This nomogram is also available online at https://surv-app.shinyapps.io/phenotype/.

Figure 3.. A comparison of C-indices for STARS vs AJCC.

STARS was validated internally (using NCDB) and externally (using SEER), and it has a statistically better fit in NCDB data than the current AJCC stage grouping (p-value from the likelihood-ratio test <0.0001).