CS372 Artificial Intelligence for Disease Diagnosis and Information Recommendations

Spring 2020             


3 units/18 sessions, including 15 lectures, and project proposal discussion and presentation.

Enrollment is limited to 30.


Edward Chang echang@cs.stanford.edu

Adjunct Professor, Computer Science, Stanford University

President, DeepQ Healthcare (areas: AI and healthcare)

Technical Advisor, SmartNews (areas: NLP and NLU)




         Artificial intelligence, specifically deep learning, stands out as one of the most transformative technologies of the past decade. AI can already outperform humans in several computer vision and natural language processing tasks.  However, we still face some of the same limitations and obstacles that led to the demise of the first AI boom phase five decades ago.  This research-oriented course will first review and reveal the limitations (e.g., iid assumption on training and testing data, voluminous training data requirement, and lacking interpretability) of some widely used AI algorithms, including convolutional neural networks (CNNs), transformers, reinforcement learning, and generative adversarial networks (GANs). To address these limitations, we will then explore topics including transfer learning for remedying data scarcity, knowledge-guided multimodal learning for improving data diversity, out of distribution generalization, attention mechanisms for enabling Interpretability, meta learning, and privacy-preserving training data management.  


The course will be taught through a combination of lecture and project sessions. Lectures on specialized AI applications (e.g., cancer/depression diagnosis and treatment, AI/VR for surgery, and health education) will feature guest speakers from academia and industry. Students will be assigned to work on an extensive project that is relevant to their fields of study (e.g., CS, Medicine, and Data Science). Projects may involve conducting literature surveys, formulating ideas, and implementing these ideas. Example project topics are but not limited to 1) knowledge guided GANs for improving training data diversity, 2) disease diagnosis via multimodal symptom checking, and 3) fake and biased news/information detection.


Level: CS senior/graduate, proficient in programming

Perquisite: Introductory course in AI, Statistics, or Machine Learning


Time and Location


·  Time: Tuesdays and Thursdays 9:00am to 10:20am

·  Location: McMurtry 360 (ART 360, next to Rodin Sculpture Garden)




·      No exams.

·      Reading: A list of papers grouped by topics will be posted by March 20th.  Each student selects two topics of interest to conduct literature surveys and presentation.

·      Term project: An extensive design and coding project.  Some example projects are:

o   Text and image multimodal learning,

o   Scene graph generated comic stripes, and

o   A News virtual assistant.


Lecture Topics (Tentative, subject to minor changes)


Segment #1 (3 sessions)

What AI Can Do and what It Cannot Do

·      Review: Deep Learning, Reinforcement Learning, and GANs (Chang)

·      Review: Word2vec, Transformers, BERT, and GPT (Chang)

·      Overview: AI Limitations, Obstacles, and Research Topics (Chang)


Segment #2 (5-6 sessions)

Revealing Limitations with Case Studies in Healthcare & News

·      Sparse Data Reinforcement Learning for Symptom Checking (Chang)

·      Transfer Learning and Multimodal Learning for Disease Diagnosis (Chang)

·      AI/VR Technologies for Brain Tumor Surgery (Kurtis, UCSF)

·      Cancer Causes Treatment, 2 lectures (Kao, Stanford)

·      Depression/Anxiety Detection and Treatment (Duvvuri, Stanford MD)


Segment #3 (4 sessions)

Advanced Methods for Overcoming Limitations

·      Privacy Regulations and DataXchange (a distributed ledger) for Compliance (Chang)

·      AI-assisted Programming and Debugging (Meijer, FB)

·      Attention-based Composition for Interpretability (TBD)

·      Successive Subspace Learning (SSL), Forward-Only Feature Learning for Robustness and Interpretability (Kuo, USC)


Segment #4 (3-4 sessions)

Tools: Virtual Assistant, HuggingFace, AutoML

·      Virtual Assistant (Giovanni, Stanford OVAL)

·      Dialogue and Query Refinement with NLP and RL (Stanford NLP group)

·      Tutorial: HuggingFace (Wolf, HuggingFace CTO) & AutoML (Deng, SmartNews)


Term Projects (3-4 sessions)

·      Proposals and Literature Surveys (Students)

·      Midterm Report (Students)

·      Final Report (Students)


Relationship to other AI courses

·       CS129, CS229 (fall, spring) Machine Learning (prerequisite)

·       CS221 (fall, spring) Artificial Intelligence (prerequisite)

·       CS230 (winter, spring) Deep learning (prerequisite)

·       CS231C (spring) Computer Vision and Image Analysis of Art (5% < overlap on CNNs and CV*)

·       CS231N (spring) CNN for Visual Recognition (5% < overlap on CNN and CV*)

·       CS235 (spring) Comp Methods for Biomed Image Analysis & Interpret (5% < overlap on CNNs and CV*)

·       CS271** (winter) Artificial intelligence in Healthcare (<5% < overlap on CNNs and CV*)

·       CS273B (fall) Deep Learning in Genomics and Biomedicine (complementary)

·       CS273C (spring) Cloud Computing for Biology and Healthcare (cloud focus, complementary)

·       CS330 (fall) Deep Multi-Task and Meta Learning (complementary)

·       CS421 (winter) Designing AI to Cultivate Human Well-being (no information available)


*5% overlap indicates the one lecture that covers CNN models and their applications to computer vision tasks.

**CS271 covers machine learning algorithms and their applications for analyzing medical images and medical records.

This research-oriented course covers advanced topics in overcoming present AI limitations and addressing scalability and privacy issues. 


Reading List