This research aims to address a key challenge in biomedical data analysis: integrating knowledge across datasets that differ in their feature spaces. I propose to develop and apply FASTER (Feature Alignment and Structured Transfer via Efficient Regularization), a novel two-step transfer learning algorithm tailored for heterogeneous biomedical data.
Research Questions: 1. How can predictive models be transferred between datasets when the available features only partially overlap? 2. Does structured feature alignment improve predictive accuracy and generalizability compared to existing transfer learning approaches? 3. What new biological and clinical insights can be gained by integrating datasets with complementary information?
Objectives: 1. Develop a robust feature space alignment strategy to map source data into a target feature space. 2. Implement a predictive transfer modeling step that leverages aligned features for improved outcome prediction. 3. Apply the framework to diverse biomedical datasets to evaluate performance and reproducibility across diseases and populations.
Scientific Rationale: Modern healthcare systems generate multifaceted data, but these data remain siloed due to differing collection protocols. Existing transfer learning methods generally assume homogeneous features, which restricts their real-world utility. FASTER directly addresses this limitation by providing a principled alignment step followed by a predictive transfer model. This approach enables meaningful integration of heterogeneous datasets, improves model generalizability, and supports the discovery of clinically relevant predictors. Ultimately, the project will contribute both a methodological advance in transfer learning and new insights into patient health through more comprehensive data integration.
