Academia Sinica Logo

Artificial intelligence is generating unprecedented opportunities for biological discovery, yet many seemingly sophisticated translational studies still fail to influence clinical care. In this lecture, I will argue that translational AI should be designed backward from the intended clinical decision rather than forward from available data. I will outline a practical framework for moving from biological insight to deployable medicine. The emphasis is not on whether a model can improve AUC in a controlled dataset, but on whether it can survive external validation, guide action, improve outcomes, and justify its cost within real healthcare systems. This perspective is directly informed by my broader research program, which focuses on clinically grounded AI, opportunistic screening from routine data, digital trial infrastructure, and real-world implementation. I will also use examples from my own work in AI-enabled electrocardiography and chest radiography to illustrate an alternative translational pathway: one that begins with latent signal discovery in routine examinations, proceeds through pragmatic randomized trials, and extends to economic evaluation, community screening, and home-based deployment. This clinical AI cycle has enabled not only large-scale model development, but also technology transfer, regulatory progress, and prospective evidence that AI can improve care when linked to pre-specified clinical action. By integrating biological depth with epidemiology-informed design, pragmatic evaluation, and deployment thinking, we can build collaborations that move beyond proof-of-concept and produce tools that are scientifically credible, clinically persuasive, and capable of changing real-world medicine.