To provide students with an understanding of the capabilities and limitations of healthcare data science sufficient to (a) design and manage research projects in this area, (b) collaborate and communicate effectively with data scientists and data science researchers, and (c) critically evaluate data science methodology, with an emphasis on model validation. Students will be exposed to the material through lectures, hands-on group exercises, written exercises drawing from the scientific literature, and a final project in data science research design.
Data science and machine learning are now beginning to impact clinical medicine, with performance on some tasks, such as detection of skin cancer, exceeding that of experienced clinicians. This course is designed to introduce students to the data science techniques poised to disrupt clinical practice through foundational material and clinical case studies. Course content will provide students with an intuitive, applications-oriented foundation in these techniques while highlighting both their capabilities and current limitations. Students will be introduced to pitfalls commonly encountered when developing models for clinical data as well as relevant practical and ethical considerations.
The course will introduce students to healthcare data science methods in the following areas:
- Logistic regression for tabular clinical data
- Neural network based prediction models, with applications to the electronic health record
- Convolutional neural networks for medical image analysis
- Natural language processing for patient notes and other text data
- Introduction to working with multi-modal health data
By the end of the course, students will understand:
- the model development process and common pitfalls
- choosing and evaluating clinically relevant performance metrics
- how to assess model generalization and possible overfitting
- the role of model interpretability and explainability in clinical decision support systems
The course will culminate in a final design project in which students will work in their teams to propose a novel data science project related to a clinical topic of their interest. Projects will focus on one of the data science areas listed above and detail the proposed approach to data collection or extraction, model development, and model validation.
- Readings will be drawn from the scientific literature. There is no course textbook.
- All readings should be accessible via the Duke network, therefore VPN in (so you have a Duke IP address) to access and download the readings from the links.
- PDFs of readings are also available in the Resources folder in Sakai
The table below lists course topics by session. For details and specific activities, please refer to the main page.
| Session | Topic |
|---|---|
| 1A | Intro to predictive models: logistic regression |
| 1B | Evaluating performance |
| 2A | The multilayer perceptron |
| 2B | The model development process |
| 3A | Medical image analysis |
| 3B | Model learning |
| 4A | Natural language processing for clinical text |
| 4B | Mitigating overfitting |
| 5A | Working with multi-modal health data |
| 5B | Understanding model predictions |
| 6A | Final project presentations |
| 6B | Final project presentations |
Class attendance is mandatory. Attendance will be taken at every class. If you miss a class, you need to make up the class session. This can be accomplished by reviewing the recorded class lecture and submitting a page synopsis of the class to the professor within one week. The one page synopsis needs to contain references. If you miss a class without prior approval by faculty, there will be a reduction in your debate/discussion overall grade.
- One missed class: 5% reduction in grade
- Two missed classes: 10% reduction in grade
- Three missed classes: 20% reduction in grade
- Four or more missed classes will result in zero grade. Class discussion is an essential element of this course. We will have written cases to discuss as well as time to review some of the major learning from team homework assignments.
| Item | In Groups? | Due Date | Percentage |
|---|---|---|---|
| Class Participation | No | Term | 20% |
| Readings and Quizzes | No | Before weekends 1-5 | 15% |
| Group Assignments | Yes | Before weekends 2, 3, 4, and 5 | 30% |
| Final Project Proposal | Yes | Before weekend 4 | 5% |
| Final Project Report | Yes | Before weekend 6 | 18% |
| Final Project Presentation | Yes | Before weekend 6 | 12% |
All assignments will be submitted through Sakai.
The MMCi Program is designed to produce the next generation of clinical informatics leaders grounded in sound ethical and scientific principles. Central among these are principles of diversity and inclusion, which will enhance our students' ability to improve human health effectively and equitably. We affirm principles of diversity and inclusion put forth by the Duke University School of Medicine and Department of Biostatistics & Bioinformatics, and we are committed to proactively fostering an inclusive environment in which diverse perspectives and backgrounds are welcome and thrive. Diverse perspectives and backgrounds are particularly critical in machine learning to ensure that decision-making algorithms are used ethically, free from bias, and reduce rather than perpetuate disparities and inequities. We are always looking for ways to make course content more inclusive, and encourage students with recommendations or concerns related to diversity and inclusion in the course content to contact the course director.
It is an expectation that everyone will be on time to class. If you are tardy for 2 or more classes without prior approval from faculty, this will also reflect in your discussion grade.
Please review the Code of Professional Conduct for the Duke University School of Medicine here
Code of Professional Conduct of the School of Medicine Preamble
"The Duke University School of Medicine strives to educate health professional students who have a high capacity for ethical professional behavior. Since training in professional behavior is a part of training in the health professions enrolled students commit themselves to comply with all regulations regarding conduct established by Duke University (the Community Standard and the Bulletin of Information and Regulations of Duke University), the School of Medicine and the individualʹs own academic program, as well as the Social Media Policy of the Duke University Health System (link). Professionalism is an academic issue and failure to demonstrate prescribed professional standards may jeopardize advancement and graduation in the same way as other academic matters. These standards closely follow those established and expected for the medical profession for which the student is training and are intended to serve as a precursor to future professional expectations."