Output based transfer learning with least squares support vector machine and its application in bladder cancer prognosis

Two dilemmas frequently occur in many real-world clinical prognoses. First, the on-hand data cannot be put entirely into the existing prediction model, since the features from new data do not perfectly match those of the model. As a result, some unique features collected from the patients in the current domain of interest might be wasted. Second, the on-hand data is not sufficient enough to learn a new prediction model. To overcome these challenges, we propose an output-based transfer learning approach with least squares support vector machine (LS-SVM) to make the maximum use of the small dataset and guarantee an enhanced generalization capability. The proposed approach can learn a current domain of interest with limited samples effectively by leveraging the knowledge from the predicted outputs of the existing model in the source domain. Also, the extent of output knowledge transfer from the source domain to the current one can be automatically and rapidly determined using a proposed fast leave-one-out cross validation strategy. The proposed approach is applied to a real-world clinical dataset to predict 5-year overall and cancer-specific mortality of bladder cancer patients after radical cystectomy. The experimental results indicate that the proposed approach achieves better classification performances than the other comparative methods and has the potential to be implemented into the real-world context to deal with small data problems in cancer prediction and prognosis.