Editing AI in Healthcare (section)

== <span style="color: #FFFFFF;">Understanding</span> ==
Healthcare AI encompasses several distinct problem types, each with its own data modalities and challenges:

'''Medical imaging''' is the most mature application. Radiology AI reads X-rays, CT scans, and MRIs — tasks that require pattern recognition similar to computer vision on natural images, but with critical differences: images are 3D (volumetric), annotations require expert radiologists, class imbalance is extreme (most scans are normal), and errors have life-or-death consequences.

'''Clinical predictive modeling''' learns from EHR data — longitudinal records of diagnoses, medications, labs, and vitals — to predict outcomes like hospital readmission, sepsis, or mortality. The challenge: EHR data is messy (missing values, inconsistent coding, free-text notes), and the temporal dynamics are complex.

'''Drug discovery''' AI accelerates the most expensive phase of pharmaceutical development. AlphaFold 2 (DeepMind, 2020) predicted protein structure from sequence with near-experimental accuracy, solving a 50-year-old problem. Generative models design novel molecules with desired properties. Graph neural networks predict molecular toxicity, solubility, and binding affinity.

'''The trust problem''' is central to healthcare AI. A radiologist needs to understand why an AI flagged a finding — not just that it did. A clinician needs to know when to trust the AI's recommendation and when to override it. This requires explainability, calibrated confidence, and human-in-the-loop design.

'''Dataset shift''' is a pervasive challenge: a model trained at one hospital may perform poorly at another due to differences in patient population, imaging equipment, protocols, and coding practices. External validation is mandatory before deployment.
</div>

<div style="background-color: #8B0000; color: #FFFFFF; padding: 20px; border-radius: 8px; margin-bottom: 15px;">