The role of deep learning‐based survival model in improving survival prediction of patients with glioblastoma

• Published in: Cancer medicine (Malden, MA) Vol. 10; no. 20; pp. 7048 - 7059
• Main Authors: Moradmand, Hajar, Aghamiri, Seyed Mahmoud Reza, Ghaderi, Reza, Emami, Hamid
• Format: Journal Article
• Language: English
• Published: United States John Wiley & Sons, Inc 01.10.2021; John Wiley and Sons Inc; Wiley
• Subjects: Adult; Age; Age Factors; Aged; Algorithms; Bayes Theorem; Bayesian analysis; Brachytherapy; Brain cancer; Brain Neoplasms - mortality; Brain Neoplasms - pathology; Brain Neoplasms - therapy; Chemoradiotherapy; Clinical Cancer Researcher; Datasets; Decision making; decision support systems; Deep Learning; DNA Modification Methylases - blood; DNA Repair Enzymes - blood; Female; Females; Genomes; Glioblastoma; Glioblastoma - mortality; Glioblastoma - pathology; Glioblastoma - therapy; Humans; hyperparameter optimization; Isocitrate dehydrogenase; Isocitrate Dehydrogenase - blood; Karnofsky Performance Status; Machine learning; Male; Mathematical models; Medical prognosis; Methylguanine; Methyltransferase; Middle Aged; Neural networks; Optimization techniques; Patients; Precision medicine; Predictive Value of Tests; Prognosis; Proportional Hazards Models; Radiation therapy; Radiotherapy - methods; Retrospective Studies; Risk factors; Sex Factors; Surgery; Survival; Survival Analysis; Tumor Suppressor Proteins - blood; Tumors; deep learning; glioblastoma; survival analysis; hyperparameter optimization; decision support systems
• ISSN: 2045-7634; 2045-7634
• DOI: 10.1002/cam4.4230

This retrospective study has been conducted to validate the performance of deep learning‐based survival models in glioblastoma (GBM) patients alongside the Cox proportional hazards model (CoxPH) and the random survival forest (RSF). Furthermore, the effect of hyperparameters optimization methods on improving the prediction accuracy of deep learning‐based survival models was investigated. Of the 305 cases, 260 GBM patients were included in our analysis based on the following criteria: demographic information (i.e., age, Karnofsky performance score, gender, and race), tumor characteristic (i.e., laterality and location), details of post‐surgical treatment (i.e., time to initiate concurrent chemoradiation therapy, standard treatment, and radiotherapy techniques), and last follow‐up time as well as the molecular markers (i.e., O‐6‐methylguanine methyltransferase and isocitrate dehydrogenase 1 status). Experimental results have demonstrated that age (Elderly > 65: hazard ratio [HR] = 1.63; 95% confidence interval [CI]: 1.213–2.18; p value = 0.001) and tumors located at multiple lobes ([HR] = 1.75; 95% [CI]: 1.177–2.61; p value = 0.006) were associated with poorer prognosis. In contrast, age (young < 40: [HR] = 0.57; 95% [CI]: 0.343–0.96; p value = 0.034) and type of radiotherapy (others include stereotactic and brachytherapy: [HR] = 0.5; 95%[CI]: 0.266–0.95; p value = 0.035) were significantly related to better prognosis. Furthermore, the proposed deep learning‐based survival model (concordance index [c‐index] = 0.823 configured by Bayesian hyperparameter optimization), outperformed the RSF (c‐index = 0.728), and the CoxPH model (c‐index = 0.713) in the training dataset. Our results show the ability of deep learning in learning a complex association of risk factors. Moreover, the remarkable performance of the deep‐learning‐based survival model could be promising to support decision‐making systems in personalized medicine for patients with GBM. This study assesses the performance of deep learning‐based survival models in glioblastoma patients alongside the Cox proportional hazards model and the random survival forests and investigates the effect of hyperparameters optimization methods on improving the prediction accuracy of deep learning‐based survival models. Our results demonstrate the outstanding ability of deep learning in learning a complex association of risk factors.