A Machine Learning Based Framework for Predicting Overall Survival in Chordoma Patients

The purpose of this research was to compare the traditional Cox proportional hazards (CoxPH) model with machine learning (ML) approaches in predicting chordoma patients' survival rates. We developed and tested three machine learning survival-models in addition to a classic CoxPH model. The dataset was split into two parts, one for training and one for validation, at random. With fivefold cross-validation and a thousand iterations of random searching, the training dataset's hyperparameters were fine-tuned. We used the Brier score, integrated Brier score, and Concordance-index(C-index) to measure the model's accuracy. To determine the accuracy, we projected 5-&10-year survival probability using the area under the AUC, calibration-curves, and receiver operating characteristic (ROC) curves. Collecting data from 724 chordoma patients, 508 were assigned to the training cohort and 216 to the validation cohort. Nine prognostic indicators were found to be significant (p < 0.05) by Cox regression. With a C-index of 0.795 and AUCs of 0.84 and 0.88 for 5- and 10-year survival, respectively, ML models outperformed the CoxPH model. DeepSurv was the most effective model. The calibration curves showed that DeepSurv's predictions and real survival rates were highly correlated. High-and low-risk groups were successfully distinguished using the DeepSurv model for risk stratification (p < 0.01). In summary DeepSurv has the finest discernment performance and tuning among ML-strategies based on time-to-event outcomes, and it is effective in chordoma prediction.