A non-invasive nomogram predicting future lung metastasis in hepatocellular carcinoma: a multi-center cohort study

Abstract

Abstract Purpose Patients with hepatocellular carcinoma at higher risk of future lung metastasis should be identified for early diagnosis and treatments. Methods From 2006 to 2016, 352 multi-center cases were retrospectively reviewed and separated into training and validation datasets. Clinical factors and radiological parameters were used to construct models through combining backward stepwise hazard models with the least absolute shrinkage and selection operator method. Discrimination and calibration of both datasets were tested. We then subdivided patients according to our model and compared their time to lung metastasis and overall survival (OS). Multivariate regression analysis was used to determine whether subgroup was an independent factor for OS. Results The best model comprised maximum diameter, fusion lesions, ascites, alpha-fetoprotein level, and regional lymph node metastasis. The area under curve for predicting one-, two-. and three-year lung metastasis free survival were 0.78, 0.92, and 0.87 (training dataset), and 0.72, 0.72, and 0.71 (validation dataset), respectively, with sufficient calibration in both datasets. Subgroups separated according to the median score of best model showed significant differences in time to lung metastasis (training, p < 0.001; validation, p = 0.002) and OS (training, p < 0.001; validation, p < 0.001). Subgroupings were significant in multivariate regression for OS in both datasets: hazard ratio (HR) = 0.435 (95% confidence interval [CI]: 0.259–0.730), p = 0.002 in the training dataset; HR = 0.341 (95% CI: 0.178–0.653), p = 0.001 in the validation dataset. Conclusion Assisted by our model, patients at high risk of future lung metastasis could be identified. For high risk population, routine chest CT should be arranged, and more combination therapies should be explored.