An Integrated Framework for Early Patent Evaluation Based on Grant and Transfer Prediction

Early patent evaluation is challenging because many commonly used value indicators become observable only after substantial time lags. This study proposes an integrated framework that combines semantic text features and structured metadata through two complementary predictors. The grant model captures examination-oriented patentability signals, and the transfer model captures diffusion-oriented adoption signals. Their predicted probabilities are combined into an interpretable InnovationIndex based on the geometric mean, so a patent receives a high score only when both proxy dimensions are reasonably strong. The study also compares this default fusion rule with arithmetic, weighted, and minimum-based alternatives to examine ranking stability. Additional analyses include transfer-side imbalance diagnostics, grant-side citation metadata ablation, and repeated-seed robustness checks. Results on EPO biomedical patents show that examiner-citation metadata provides strong information for grant prediction, while transfer prediction depends more on contextual metadata and requires careful interpretation under class imbalance. Overall, the framework offers a practical screening tool for early-stage patent assessment when long-horizon indicators such as citations, legal events, and market outcomes are still unavailable.