Stanford AI Course Map 中文 GitHub ↗
← Back to the course map
Stanford · 2026—27 CS 287 AI for Science & Health ★★★☆☆

Core course · AI for Science & Health

Data Centric AI for Healthcare (BMDS 218)

中文:医疗 Data-Centric AI

Data-centric AI for healthcare: how data is generated, how bias/missingness/leakage break models, and how to fix it.

Draft translation: official facts are checked; wording may change after native review.

Offered in Winter 2026–27; no stable course-site link found.

3 units Winter 2027 Medical Option (Med-Ltr-CR/NC) ★★★☆☆
HealthcareDataData-centricRobustness

Course positioning

Taught by Emily Alsentzer, Jason Fries, and others, CS 287 takes the opposite view from most ML courses: it studies data, not models. It discusses how clinical workflows, documentation habits, and institutional practices introduce bias, noise, missingness, and spurious correlations, and gives practical methods for data quality, labeling, synthetic data, and drift monitoring.

What it covers

  1. Data generation: clinical workflows, documentation, institutions
  2. Data quality: bias, noise, missingness, spurious correlations
  3. Label leakage: detection and avoidance
  4. Labeling strategies: scalable, accurate annotation
  5. Synthetic data: privacy, augmentation, risks
  6. Drift and monitoring: data/model drift
  7. Generalization and robustness
  8. Cases and project

Prerequisites

Machine learning basics; interest in healthcare data.

Assessment & workload

Case discussion plus a project/assignment. Check the current syllabus for details.

Who it's for

Medical AI, clinical data science, and data engineering; anyone frustrated by 'the model is great but breaks in production'.

Study advice

It complements CS 277: 277 is models, 287 is data. Try to touch a real clinical data workflow for the project.

Related courses

Official course description

This course explores how healthcare data is generated in practice - and how these processes influence the development of robust, trustworthy machine learning models. Students will examine how clinical workflows, documentation habits, and institutional practices introduce bias, noise, missingness, and spurious correlations into datasets. The course emphasizes the often-overlooked reality that many model failures stem not from algorithmic design, but from subtle flaws in how data is collected, labeled, and interpreted. Through lectures, real-world case studies, and observation of clinical data workflows, students will develop practical tools for understanding and improving data quality and model reliability. Topics include assessing data quality, addressing missingness and label leakage, developing scalable and accurate annotation strategies, leveraging synthetic data, and monitoring data and models for drift over time. Students will also learn to evaluate how these data challenges impact model generalization and robustness.