Novel artificial intelligence machine learning approaches to precisely predict survival and site-specific recurrence in cervical cancer: A multi-institutional study

BackgroundMachine learning (ML) has been gradually integrated into oncologic research but seldom applied to predict cervical cancer (CC), and no model has been reported to predict survival and site-specific recurrence simultaneously. Thus, we aimed to develop ML models to predict survival and site-specific recurrence in CC and to guide individual surveillance.MethodsWe retrospectively collected data on CC patients from 2006 to 2017 in four hospitals. The survival or recurrence predictive value of the variables was analyzed using multivariate Cox, principal component, and K-means clustering analyses. The predictive performances of eight ML models were compared with logistic or Cox models. A novel web-based predictive calculator was developed based on the ML algorithms.ResultsThis study included 5112 women for analysis (268 deaths, 343 recurrences): (1) For site-specific recurrence, larger tumor size was associated with local recurrence, while positive lymph nodes were associated with distant recurrence. (2) The ML models exhibited better prognostic predictive performance than traditional models. (3) The ML models were superior to traditional models when multiple variables were used. (4) A novel predictive web-based calculator was developed and externally validated to predict survival and site-specific recurrence.ConclusionML models might be a better analytic approach in CC prognostic prediction than traditional models as they can predict survival and site-specific recurrence simultaneously, especially when using multiple variables. Moreover, our novel web-based calculator may provide clinicians with useful information and help them make individual postoperative follow-up plans and further treatment strategies.     

Identification of Plasma Lipid Biomarkers for Prostate Cancer by Lipidomics and Bioinformatics

Background

Lipids have critical functions in cellular energy storage, structure and signaling. Many individual lipid molecules have been associated with the evolution of prostate cancer; however, none of them has been approved to be used as a biomarker. The aim of this study is to identify lipid molecules from hundreds plasma apparent lipid species as biomarkers for diagnosis of prostate cancer.

Methodology/Principal Findings

Using lipidomics, lipid profiling of 390 individual apparent lipid species was performed on 141 plasma samples from 105 patients with prostate cancer and 36 male controls. High throughput data generated from lipidomics were analyzed using bioinformatic and statistical methods. From 390 apparent lipid species, 35 species were demonstrated to have potential in differentiation of prostate cancer. Within the 35 species, 12 were identified as individual plasma lipid biomarkers for diagnosis of prostate cancer with a sensitivity above 80%, specificity above 50% and accuracy above 80%. Using top 15 of 35 potential biomarkers together increased predictive power dramatically in diagnosis of prostate cancer with a sensitivity of 93.6%, specificity of 90.1% and accuracy of 97.3%. Principal component analysis (PCA) and hierarchical clustering analysis (HCA) demonstrated that patient and control populations were visually separated by identified lipid biomarkers. RandomForest and 10-fold cross validation analyses demonstrated that the identified lipid biomarkers were able to predict unknown populations accurately, and this was not influenced by patient''s age and race. Three out of 13 lipid classes, phosphatidylethanolamine (PE), ether-linked phosphatidylethanolamine (ePE) and ether-linked phosphatidylcholine (ePC) could be considered as biomarkers in diagnosis of prostate cancer.

Conclusions/Significance

Using lipidomics and bioinformatic and statistical methods, we have identified a few out of hundreds plasma apparent lipid molecular species as biomarkers for diagnosis of prostate cancer with a high sensitivity, specificity and accuracy.     

Tissue biomarkers for prostate cancer radiation therapy

Prostate cancer is the most common cancer and second leading cause of cancer deaths among men in the United States. Most men have localized disease diagnosed following an elevated serum prostate specific antigen test for cancer screening purposes. Standard treatment options consist of surgery or definitive radiation therapy directed by clinical factors that are organized into risk stratification groups. Current clinical risk stratification systems are still insufficient to differentiate lethal from indolent disease. Similarly, a subset of men in poor risk groups need to be identified for more aggressive treatment and enrollment into clinical trials. Furthermore, these clinical tools are very limited in revealing information about the biologic pathways driving these different disease phenotypes and do not offer insights for novel treatments which are needed in men with poor-risk disease. We believe molecular biomarkers may serve to bridge these inadequacies of traditional clinical factors opening the door for personalized treatment approaches that would allow tailoring of treatment options to maximize therapeutic outcome. We review the current state of prognostic and predictive tissue-based molecular biomarkers which can be used to direct localized prostate cancer treatment decisions, specifically those implicated with definitive and salvage radiation therapy.     

Proteomic analysis of conditioned media from the PC3, LNCaP, and 22Rv1 prostate cancer cell lines: discovery and validation of candidate prostate cancer biomarkers

Early detection of prostate cancer is problematic due to the lack of a marker that has high diagnostic sensitivity and specificity. The prostate specific antigen (PSA) test, in combination with digital rectal examination, is the gold standard for prostate cancer diagnosis. However, this modality suffers from low specificity. Therefore, specific markers for clinically relevant prostate cancer are needed. Our objective was to proteomically characterize the conditioned media from three human prostate cancer cell lines of differing origin [PC3 (bone metastasis), LNCaP (lymph node metastasis), and 22Rv1 (localized to prostate)] to identify secreted proteins that could serve as novel prostate cancer biomarkers. Each cell line was cultured in triplicate, followed by a bottom-up analysis of the peptides by two-dimensional chromatography and tandem mass spectrometry. Approximately, 12% (329) of the proteins identified were classified as extracellular and 18% (504) as membrane-bound among which were known prostate cancer biomarkers such as PSA and KLK2. To select the most promising candidates for further investigation, tissue specificity, biological function, disease association based on literature searches, and comparison of protein overlap with the proteome of seminal plasma and serum were examined. On the basis of this, four novel candidates, follistatin, chemokine (C-X-C motif) ligand 16, pentraxin 3 and spondin 2, were validated in the serum of patients with and without prostate cancer. The proteins presented in this study represent a comprehensive sampling of the secreted and shed proteins expressed by prostate cancer cells, which may be useful as diagnostic, prognostic or predictive serological markers for prostate cancer.     

Semiparametric Models of Time-Dependent Predictive Values of Prognostic Biomarkers

Summary .  Rigorous statistical evaluation of the predictive values of novel biomarkers is critical prior to applying novel biomarkers into routine standard care. It is important to identify factors that influence the performance of a biomarker in order to determine the optimal conditions for test performance. We propose a covariate-specific time-dependent positive predictive values curve to quantify the predictive accuracy of a prognostic marker measured on a continuous scale and with censored failure time outcome. The covariate effect is accommodated with a semiparametric regression model framework. In particular, we adopt a smoothed survival time regression technique ( Dabrowska, 1997 ,  The Annals of Statistics   25, 1510–1540) to account for the situation where risk for the disease occurrence and progression is likely to change over time. In addition, we provide asymptotic distribution theory and resampling-based procedures for making statistical inference on the covariate-specific positive predictive values. We illustrate our approach with numerical studies and a dataset from a prostate cancer study.     

Strategies to develop radiomics and machine learning models for lung cancer stage and histology prediction using small data samples

Predictive models based on radiomics and machine-learning (ML) need large and annotated datasets for training, often difficult to collect. We designed an operative pipeline for model training to exploit data already available to the scientific community. The aim of this work was to explore the capability of radiomic features in predicting tumor histology and stage in patients with non-small cell lung cancer (NSCLC).We analyzed the radiotherapy planning thoracic CT scans of a proprietary sample of 47 subjects (L-RT) and integrated this dataset with a publicly available set of 130 patients from the MAASTRO NSCLC collection (Lung1). We implemented intra- and inter-sample cross-validation strategies (CV) for evaluating the ML predictive model performances with not so large datasets.We carried out two classification tasks: histology classification (3 classes) and overall stage classification (two classes: stage I and II). In the first task, the best performance was obtained by a Random Forest classifier, once the analysis has been restricted to stage I and II tumors of the Lung1 and L-RT merged dataset (AUC = 0.72 ± 0.11). For the overall stage classification, the best results were obtained when training on Lung1 and testing of L-RT dataset (AUC = 0.72 ± 0.04 for Random Forest and AUC = 0.84 ± 0.03 for linear-kernel Support Vector Machine).According to the classification task to be accomplished and to the heterogeneity of the available dataset(s), different CV strategies have to be explored and compared to make a robust assessment of the potential of a predictive model based on radiomics and ML.     

A data-analytic strategy for protein biomarker discovery: profiling of high-dimensional proteomic data for cancer detection

With recent advances in mass spectrometry techniques, it is now possible to investigate proteins over a wide range of molecular weights in small biological specimens. This advance has generated data-analytic challenges in proteomics, similar to those created by microarray technologies in genetics, namely, discovery of 'signature' protein profiles specific to each pathologic state (e.g. normal vs. cancer) or differential profiles between experimental conditions (e.g. treated by a drug of interest vs. untreated) from high-dimensional data. We propose a data-analytic strategy for discovering protein biomarkers based on such high-dimensional mass spectrometry data. A real biomarker-discovery project on prostate cancer is taken as a concrete example throughout the paper: the project aims to identify proteins in serum that distinguish cancer, benign hyperplasia, and normal states of prostate using the Surface Enhanced Laser Desorption/Ionization (SELDI) technology, a recently developed mass spectrometry technique. Our data-analytic strategy takes properties of the SELDI mass spectrometer into account: the SELDI output of a specimen contains about 48,000 (x, y) points where x is the protein mass divided by the number of charges introduced by ionization and y is the protein intensity of the corresponding mass per charge value, x, in that specimen. Given high coefficients of variation and other characteristics of protein intensity measures (y values), we reduce the measures of protein intensities to a set of binary variables that indicate peaks in the y-axis direction in the nearest neighborhoods of each mass per charge point in the x-axis direction. We then account for a shifting (measurement error) problem of the x-axis in SELDI output. After this pre-analysis processing of data, we combine the binary predictors to generate classification rules for cancer, benign hyperplasia, and normal states of prostate. Our approach is to apply the boosting algorithm to select binary predictors and construct a summary classifier. We empirically evaluate sensitivity and specificity of the resulting summary classifiers with a test dataset that is independent from the training dataset used to construct the summary classifiers. The proposed method performed nearly perfectly in distinguishing cancer and benign hyperplasia from normal. In the classification of cancer vs. benign hyperplasia, however, an appreciable proportion of the benign specimens were classified incorrectly as cancer. We discuss practical issues associated with our proposed approach to the analysis of SELDI output and its application in cancer biomarker discovery.     

Performance of an adipokine pathway-based multilocus genetic risk score for prostate cancer risk prediction

Few biomarkers are available to predict prostate cancer risk. Single nucleotide polymorphisms (SNPs) tend to have weak individual effects but, in combination, they have stronger predictive value. Adipokine pathways have been implicated in the pathogenesis. We used a candidate pathway approach to investigate 29 functional SNPs in key genes from relevant adipokine pathways in a sample of 1006 men eligible for prostate biopsy. We used stepwise multivariate logistic regression and bootstrapping to develop a multilocus genetic risk score by weighting each risk SNP empirically based on its association with disease. Seven common functional polymorphisms were associated with overall and high-grade prostate cancer (Gleason≥7), whereas three variants were associated with high metastatic-risk prostate cancer (PSA≥20 ng/mL and/or Gleason≥8). The addition of genetic variants to age and PSA improved the predictive accuracy for overall and high-grade prostate cancer, using either the area under the receiver-operating characteristics curves (P<0.02), the net reclassification improvement (P<0.001) and integrated discrimination improvement (P<0.001) measures. These results suggest that functional polymorphisms in adipokine pathways may act individually and cumulatively to affect risk and severity of prostate cancer, supporting the influence of adipokine pathways in the pathogenesis of prostate cancer. Use of such adipokine multilocus genetic risk score can enhance the predictive value of PSA and age in estimating absolute risk, which supports further evaluation of its clinical significance.     

Patient-specific IMRT QA verification using machine learning and gamma radiomics

Gamma function is the standard methodology for comparing dose distributions. It is calculated in dedicated software, and its results verification is not performed. Thus we developed an automatic tool for patient-specific QA results verification through high accuracy machine learning (ML) models based on the radiomics characteristics extraction from gamma images. We used 158 patient-specific QA tests and extracted 105 radiomics features from each gamma image. Three random forest models were developed (ML I, ML II, and ML III). ML I and ML II verified the gamma image approval using criteria of 2%/2mm/15% threshold and 3%/3mm/15% threshold, respectively. ML III verified if the gamma analyzes software recommended protocol was followed to detect if the TPS grid modification step was done. The models were based on the most important features selected using the mean decreased impurity, and their performances were evaluated. ML I included 25 features. Its accuracy was 0.85 using the test set and 0.84 using dataset B. ML II included 10 features, and its accuracy with the test set was 0.98; the same value was achieved using the never seen data (dataset B). The First-order 10th percentile feature was identified as a feature strongly related to the approved classification. ML III selected 23 features with an accuracy of 0.99 for test set and 0.98 for dataset B. An automatic workflow example for gamma analyses QA results verification could be proposed combining the models to detect grid inconsistencies on software evaluation, followed by the test approval classification.     

Quantitative Analysis of BTF3, HINT1, NDRG1 and ODC1 Protein Over-Expression in Human Prostate Cancer Tissue

Prostate carcinoma is the most common cancer in men with few, quantifiable, biomarkers. Prostate cancer biomarker discovery has been hampered due to subjective analysis of protein expression in tissue sections. An unbiased, quantitative immunohistochemical approach provided here, for the diagnosis and stratification of prostate cancer could overcome this problem. Antibodies against four proteins BTF3, HINT1, NDRG1 and ODC1 were used in a prostate tissue array (> 500 individual tissue cores from 82 patients, 41 case pairs matched with one patient in each pair had biochemical recurrence). Protein expression, quantified in an unbiased manner using an automated analysis protocol in ImageJ software, was increased in malignant vs non-malignant prostate (by 2-2.5 fold, p<0.0001). Operating characteristics indicate sensitivity in the range of 0.68 to 0.74; combination of markers in a logistic regression model demonstrates further improvement in diagnostic power. Triple-labeled immunofluorescence (BTF3, HINT1 and NDRG1) in tissue array showed a significant (p<0.02) change in co-localization coefficients for BTF3 and NDRG1 co-expression in biochemical relapse vs non-relapse cancer epithelium. BTF3, HINT1, NDRG1 and ODC1 could be developed as epithelial specific biomarkers for tissue based diagnosis and stratification of prostate cancer.     

Newer potential biomarkers in prostate cancer

Prostate-specific antigen (PSA) screening has led to a significant rise in the number of men diagnosed with prostate cancer and an associated increase in biopsies performed. Despite its limitations, including a positive predictive value of only 25%-40%, PSA remains the only generally accepted biomarker for prostate cancer. There is a need for better tools to not only identify men with prostate cancer, but also to recognize those with potentially lethal disease who will benefit from intervention. A great deal of work has been done worldwide to improve our knowledge of the genetics behind prostate cancer and the specificity of PSA by developing assays for different PSA isoforms. Common genetic alterations in prostate cancer patients have been identified, including CpG hypermethylation of GSPT1 and TMPRSS2:ERG gene fusion. Serum and urine detection of RNA biomarkers (eg, PCA3) and prostate cancer tissue protein antibodies (eg, EPCA) are being evaluated for detection and prognostic tools. This article reviews some of the promising developments in biomarkers.     

Levels of plasma glycan-binding auto-IgG biomarkers improve the accuracy of prostate cancer diagnosis

Strategies to improve the early diagnosis of prostate cancer will provide opportunities for earlier intervention. The blood-based prostate-specific antigen (PSA) assay is widely used for prostate cancer diagnosis but specificity of the assay is not satisfactory. An algorithm based on serum levels of PSA combined with other serum biomarkers may significantly improve prostate cancer diagnosis. Plasma glycan-binding IgG/IgM studies suggested that glycan patterns differ between normal and tumor cells. We hypothesize that in prostate cancer glycoproteins or glycolipids are secreted from tumor tissues into the blood and induce auto-immunoglobulin (Ig) production. A 24-glycan microarray and a 5-glycan subarray were developed using plasma samples obtained from 35 prostate cancer patients and 54 healthy subjects to identify glycan-binding auto-IgGs. Neu5Acα2-8Neu5Acα2-8Neu5Acα (G81)-binding auto-IgG was higher in prostate cancer samples and, when levels of G81-binding auto-IgG and growth differentiation factor-15 (GDF-15 or NAG-1) were combined with levels of PSA, the prediction rate of prostate cancer increased from 78.2% to 86.2% than with PSA levels alone. The G81 glycan-binding auto-IgG fraction was isolated from plasma samples using G81 glycan-affinity chromatography and identified by N-terminal sequencing of the 50 kDa heavy chain variable region of the IgG. G81 glycan-binding 25 kDa fibroblast growth factor-1 (FGF1) fragment was also identified by N-terminal sequencing. Our results demonstrated that a multiplex diagnostic combining G81 glycan-binding auto-IgG, GDF-15/NAG-1 and PSA (≥?2.1 ng PSA/ml for cancer) increased the specificity of prostate cancer diagnosis by 8%. The multiplex assessment could improve the early diagnosis of prostate cancer thereby allowing the prompt delivery of prostate cancer treatment.

     

Transrectal sonography in prostate cancer detection--our 25 years experience of implementation

Prostate cancer is a leading public health problem of male population in developed countries. Gold standard for prostate cancer diagnosis is true cut biopsy guided by transrectal ultrasound. Aim of this study was to determine sensitivity, specificity, accuracy, positive and negative predictive value of transrectal sonography (TRUS) in prostate cancer detection. The analysis was made for two time periods, before and after routine implementation of prostate specific antigen (PSA) in prostate cancer diagnostics. From 1984 to 1993 TRUS guided prostate biopsy was performed in 564, and from 1994 to 2008 in 5678 patients. In the second period PSA was routinely used in prostate cancer diagnostics. In the first period by TRUS we have made an exact diagnosis of prostate cancer in 18.97% of patients what was confirmed by biopsy. 4.61% ware false positive and 11.34% ware false negative. In the second period prostate cancer was recognized in 30.34% of patients, confirmed by biopsy. False positive cases ware 6.11% and false negative 29.31%. Sensitivity of transrectal sonography in the first period was 62.57%, specificity 94.2%, accuracy 86.2%, positive predictive value 80.45% and negative predictive value 87.72%. In the second period sensitivity was 50.87%, specificity 91.93%, accuracy 73.84%, positive predictive value 83.24% and negative predictive value 70.39%. Based on our experience we can conclude that prostate cancer is mostly found in the peripheral zone. Smaller tumors are hypoechoic and bigger tumors are hyperechoic. Prostate cancer lesions are impossible to differentiate from chronic prostatitis only by TRUS. Implementation of PSA has significantly decrease sensitivity, accuracy and negative predictive value of TRUS in prostate cancer detection. TRUS guided true cut biopsy is a gold standard in prostate cancer diagnostics.     

Identifying potential DNA methylation markers in early-stage colorectal Cancer

Colorectal cancer (CRC) is the second leading malignancy worldwide. Accurate screening is pivotal to early CRC detection, yet current screening modality involves invasive colonoscopy while non-invasive FIT tests have limited sensitivity. We applied a DNA methylation assay to identify biomarkers for early-stage CRC detection, risk stratification and precancerous lesion screening at tissue level. A model of biomarkers SFMBT2, ITGA4, THBD and ZNF304 showed 96.1% sensitivity and 87.0% specificity in CRC detection, with 100.0% sensitivity for advanced precancerous lesion and stage I CRC. Performances were further validated with TCGA data set, which showed a consistent AUC of 0.99 and exhibited specificity against other cancer types. KCNJ12, VAV3-AS1 and EVC were further identified for stage stratification (stage 0-I versus stage II-IV), with AUC of 0.87, 83.0% sensitivity and 71.2% specificity. Additionally, dual markers of NEUROD1 and FAM72C showed 83.2% sensitivity and 77.4% specificity in differing non-advanced precancerous lesions from inflammatory bowel diseases.     

Scope and limitations of MALDI-TOF MS blood serum peptide profiling in cancer diagnostics

Serum samples (33 healthy women, 34 ovarian cancer, 28 colorectal cancer, 34 syphilis patients and 136 patients with various benign gynecological diseases) were analyzed by MALDI-TOF MS peptide profiling and respective predictive models were generated by genetic and supervised neural network algorithms. Classification models for pathology versus healthy control showed up to 100% sensitivity and specificity for all target diseases. However, the specificity dropped to unsatisfactory 25–40% in case of target versus nontarget disease diagnostics. Expansion of the control group to an artificial “nominal control” group by adding profiles of benign gynecological diseases considerably improved specificity of the models distinguishing ovarian cancer from healthy control and benign gynecological diseases. The suggested version of MALDI-TOF MS profiling of sera could be applied to differentiate between cancers and benign neoplasms of the same localization which is a challenging task for classical methods. To increase the specificity of diagnostic methods based on peptidome analysis of blood samples, it is necessary to identify sets of concrete peptide structures which qualitatively or quantitatively differ among patients with different diseases.     

Machine Learning Based Computer Aided Diagnosis of Breast Cancer Utilizing Anthropometric and Clinical Features

Breast cancer is one of the most prevalent types of cancers in females, which has become rampant all over the world in recent years. The survival rate of breast cancer patients degrades considerably for patients diagnosed at an advanced stage compared to those diagnosed at an early stage. The objective of this study is two folds. The first one is to find the most relevant biomarkers of breast cancer, which can be attained from regular blood analysis and anthropometric measurements. The other one is to improve the performance of current computer-aided diagnosis (CAD) system of early breast cancer detection. This study utilized a recent data set containing nine anthropometric and clinical attributes. In our methodology, first, we performed multicollinearity analysis and ranked the features based on the weighted average score obtained from four filter-based feature evaluation methods such as F-score, information gain, chi-square statistic, and Minimum Redundancy Maximum Relevance. Next, to improve the separability of the target classes, we scaled and weighted the dataset using min-max normalization and similarity-based attribute weighting by the k-means clustering algorithm, respectively. Finally, we trained standard machine learning (ML) models and evaluated the performance metrics by 10-fold cross-validation method. Our support vector machine (SVM) model with radial basis function (RBF) kernel appeared to be the most successful classifier by utilizing six features, namely, Body Mass Index (BMI), Age, Glucose, MCP-1, Resistin, and Insulin. The obtained classification accuracy, sensitivity, and specificity are 93.9% (95% CI: 93.2–94.6%), 95.1% (95% CI: 94.4–95.8%), and 94.0% (95% CI: 93.3–94.7%), respectively; these performance metrics outperformed state-of-the-art methods reported in the literature. The developed model could potentially assist the medical experts for the early diagnosis of breast cancer by employing a set of attributes that can be easily obtained from regular blood analysis and anthropometric measurements.     

Generation of “Virtual” Control Groups for Single Arm Prostate Cancer Adjuvant Trials

It is difficult to construct a control group for trials of adjuvant therapy (Rx) of prostate cancer after radical prostatectomy (RP) due to ethical issues and patient acceptance. We utilized 8 curve-fitting models to estimate the time to 60%, 65%, … 95% chance of progression free survival (PFS) based on the data derived from Kattan post-RP nomogram. The 8 models were systematically applied to a training set of 153 post-RP cases without adjuvant Rx to develop 8 subsets of cases (reference case sets) whose observed PFS times were most accurately predicted by each model. To prepare a virtual control group for a single-arm adjuvant Rx trial, we first select the optimal model for the trial cases based on the minimum weighted Euclidean distance between the trial case set and the reference case set in terms of clinical features, and then compare the virtual PFS times calculated by the optimum model with the observed PFSs of the trial cases by the logrank test. The method was validated using an independent dataset of 155 post-RP patients without adjuvant Rx. We then applied the method to patients on a Phase II trial of adjuvant chemo-hormonal Rx post RP, which indicated that the adjuvant Rx is highly effective in prolonging PFS after RP in patients at high risk for prostate cancer recurrence. The method can accurately generate control groups for single-arm, post-RP adjuvant Rx trials for prostate cancer, facilitating development of new therapeutic strategies.     

A novel seven-lncRNA signature for prognosis prediction in hepatocellular carcinoma

Liver cancer is still one of the leading causes of cancer-related death worldwide. This study is dedicated to developing a multi–long noncoding RNA (lncRNA) model for risk stratification and prognosis prediction on patients with hepatocellular carcinoma (HCC). We first downloaded lncRNA expression profiles and corresponding clinical information of patients with liver cancer from The Cancer Genome Atlas database. Differentially expressed (DE) lncRNAs between HCC samples and normal samples were identified. In total, 308 patients with HCC were randomly divided into a training group (n = 154) and a testing group (n = 154). Univariate Cox regression and least absolute shrinkage and selection operator Cox regression analyses were performed to select the best survival-related candidates from these DE lncRNAs in the training set. Seven lncRNAs (AC009005.2, RP11-363N22.3, RP11-932O9.10, RP11-572O6.1, RP11-190C22.8, RP11-388C12.8, and ZFPM2-AS1) were finally identified and used to construct a seven-lncRNA signature. The signature could classify patients into high-risk and low-risk groups with significantly different overall survival. The area under the curve of receiver operating characteristic curve for the signature to predict 5-year survival reached more than 0.75. Besides, the prognostic value of the seven-lncRNA signature was independent of conventional clinical factors. The predictive performance of the signature was further validated in the testing set and the whole set. Functional enrichment analysis indicated that the seven prognostic lncRNAs may be involved in several essential biological processes and pathways. The current study demonstrated the potential clinical implications of the seven-lncRNA signature for survival prediction of patients with HCC.