Deep-learning based 3-year survival prediction of pineoblastoma patients

Abstract

Abstract Purpose Pineoblastoma (PB) is an extremely uncommon and highly aggressive malignancy that originates from the pineal gland, more frequently occurs in young children. Due to the rare nature, little is known about its prognostic implications and survival outcomes. Existing methods for prognostication based on traditional statistical approaches such as Cox proportional hazards (CPH) models, which have less-than-ideal predictive accuracy. Recently, deep learning algorithms has unlocked unprecedented advancements in diverse domains and has been applied extensively in medical fields. Thus, we sought to develop and compare deep learning models with CPH models in predicting 3-year overall (OS) and disease-specific survival (DSS) for patients with pineoblastoma. Methods We utilized the Surveillance, Epidemiology, and End Results (SEER) database to identify patients diagnosed with pineoblastoma between 1975 and 2019. The dataset divided into training and testing sets (70:30 split) for training and evaluating deep neural networks (DNN) models, while 5-fold cross-validation was employed. Additionlly, multivariable CPH models were established for comparison. The primary endpoint was 3-year overall survival (OS) and disease-specific survival (DSS). The performance of the models was evaluated using the area under the receiver operating characteristic curve (AUC) and calibration curve. Results A total of 145 patients were included in the study. The AUC value for the DNN models was 0.92 for OS and 0.91 for DSS. In comparison, the AUC value for the CPH models was 0.641 for OS and 0.685 for DSS. Meanwhile, the DNN models demonstrated good calibration: OS model (slope = 0.94, intercept = 0.07) and DSS model (slope = 0.81, intercept = 0.20). Conclusions The DNN models that we constructed exhibited excellent predictive capabilities in forecasting the 3-year survival of pineoblastoma patients, outperforming the CPH models. Deep learning is expected to aid clinicians predict the prognosis effectively and accurately for patients with rare tumors.