To the best of our knowledge, this is the first published study that uses machine learning algorithms to build prediction models for the survival of pineoblastoma patients. We demonstrated that deep network models have excellent and powerful capabilities in prognostication, and outperformed traditional Cox proportional hazards models to predict 3-year OS and DSS. Furthermore, Multivariate Cox model analysis suggested that age and chemotherapy were the independent protective factors, while distant metastasis was an independent risk factor for all-cause mortality.
In this study, our DNN models have exhibited superior predictive power and achieved an AUC of 0.92 for overall and 0.91 for disease-specific survival. Moreover, the DNN models were well-calibrated, with a near-perfect calibration slope of 0.94 and an intercept of 0.07. Overall, the DNN models have promising discriminatory accuracy and good calibration performance, and are significantly better than CPH models. It also suggested that the DNN models can be used as an accurate and reliable predictor of 3-year survival in PB patients. Compared with traditional statistical methods, the main advantage of deep learning can extract linear and non-linear relationships between clinical variables and survival outcomes through a multi-layer network structure and a non-linear activation function[23]. This advantage likely renders our DNN models superior to CPH models in the survival prediction of patients with pineoblastoma. In several previous studies, most of the survival prediction models (i.e Cox proportional hazards model, Nomograms) have been developed focus on either pediatric pineoblastoma or adults pineoblastoma, and the best AUC value only reached 0.802[14]. However, we constructed DNN models based on pineoblastoma patients of all ages and achieved a near-perfect AUC value. The capacity to detect non-linear association among covariates might be the underlying cause of our DNN models compared favorably with other models in the literature.
Although deep learning models confer good discriminative power and high predictive accuracy, there is a common criticism of deep learning that these models have limited interpretability of the weighting and interactions of the variables due to their black-boxes nature. In this line, CPH models offer more information regarding the prognosis and have better interpretability in terms of the effect of clinical variables on survival outcomes. In the present study, using CPH models, we obtained the hazard ratios of every variable and determined increasing age, distant metastatic, and chemotherapy as independent prognostic factors for overall survival. In contrast, DNN models did not possess this extent of explanatory power.
Recently, with the explosive development of artificial intelligence, machine learning, especially deep learning, is progressing and expanding rapidly, and has been intensively adopted in many medical domains. Furthermore, the feasibility and utility of deep learning algorithms have been shown in various studies. Using DL methods, Sebastian et al. developed a multi-task 3D-CNN model (the convolutional neural network that uses the full 3D) based on multimodal MRI features to predict the subtype of gliomas and achieved a high performance (AUC 0.90)[24]. Likewise, a study by Bangalore et al. used an automated deep-learning-based network to detect IDH mutation status noninvasively based on T2-MRI images of patients with gliomas, with accuracies of 97.14% and sensitivity of 97%[25]. A deep learning image signature model conducted by Yan et al. study, showing an AUC value of 0.999 for the training dataset, 0.986 for the validation dataset, and 0.983 for the testing dataset in predicting the 1p/19q codeletion status in patients with low-grade gliomas[26]. Other studies also have confirmed the ability of DL to identify EGFR amplification status[27, 28], MGMT hypermethylation status[29]. Beyond this, a great promise of deep learning methods in prognosticating overall survival is one that has been displayed in dozens of studies as well. Tang et al. Study shown that a multi-task convolutional neural network (CNN) model can accomplish OS prediction based on pre-operative MRI images in glioblastoma patients and achieved a high accuracy (AUC 0.946)[30]. Sun et al. also proposed an automated DL-based prognostic models for spinal cord astrocytoma patients, of which its AUC value in predicting 1-year, 3-year, 5-year OS was 0.881 (95% CI 0.839–0.918), 0.862 (95% CI 0.827–0.901), and 0.905 (95% CI 0.867–0.942), respectively[31]. This suggested that deep learning is emerging as a potentially powerful tool for predicting oncologic events.
Owing to the rareness of pineoblastoma, large-scale, specific trials can be difficult to conduct. Most studies of this tumor are single-center, retrospective analyses with small sample sizes. In these studies, the associations between relevant prognostic features and survival outcomes have not been robustly identified due to insufficient data and a lack of generalizability. Recently, the growing presence and increasing availability of large databases present an excellent opportunity to explore rare tumors. SEER database is a publicly accessible, extensive, and longstanding repository that encompasses comprehensive clinical information of a substantial cohort of cancer patients in the United States. It provides an invaluable opportunity to access population-level data on patients with pineoblastoma, enabling the development of generalizable deep neural network (DNN) models and the identification of underlying prognostic factors.
However, it is important to acknowledge the limitations of our study. Firstly, our DNN models have not been validated in an external data set, and rigorous external validation is necessary to confirm these initial findings and to determine whether a model applies to other study cohorts. While external validation was not conducted in the models developed within this study, the utilization of a nationally curated database, encompassing data from 48% of the United States, is expected to enhance the generalizability of findings. Moreover, in order to include as many cases as possible, we have assembled clinical data comprehensively on all patients with pineoblastoma diagnosed between 1975 to 2019. Second, the SEER database only provides relatively limited information, lacking complete and detailed treatment data, including radiation dosage, radiation route, and chemotherapy regimen. we, therefore, are unable to further comment on the effects of the above factors on survival outcomes. Despite these limitations, our DNN models still show robust predictive performance in 3-year survival of patients with pineoblastoma. It was implied that deep learning has a considerable potential for helping clinicians predict the prognosis effectively and accurately based on the individual clinical features. Finally, an inherent difficulty in the study of pineoblastoma is the rarity of this disease and this is also reflected in our study.