biomedical data science course: brainstorming

[SaraMati]

to follow up the idea that I raised with you a while back on designing a graduate course. Please pour in your ideas and whether you want to be part of it. This is a very rough draft:

Biomedical data science

Target population

Graduate students that are from different backgrounds and have entered the Neuromodulation/Neuroscience/biomedical engineering program

Course Description

Biomedical data includes a diverse range of data types from patient data from national registeries, to biomarker data, to genetic and sequence data, to electrophysiological recordings from human or animal models, to simulations; usually covering multiple spatio-temporal scales. These data are usually high dimensional, noisy and fragmented. Handling such data is challenging and understanding it is the very important first step in a data-driven approach to build predictive models. This course introduces techniques to analyze biomedical data using programing languages Python and R. The focus of the course is to integrate reproducible research principles such as well-documented modular coding, correctness of procedure, and interpretability of results and presentation of outcomes of the analysis.

Learning Outcomes

• understand complications related to using and analyzing biomedical data; with a focus in high dimensional data
• write well written, well documented modular code
• manipulate, clean and filter data and organize data properly
• discriminate between interpretable and black box models and correctly setup validated statistical models
• generate publication quality presentation of outcomes and effectively communicate technical content
• productive collaborative work in a diverse team

Ideas for course syllabus

Programming

• Intro to Python/R
• Data cleaning/wrangling - tidy data
• From Data Exploration to Publication-Quality Figures
• scientific computing skills: version control and GitHub; documentation; modular coding; Quantitative methods
• best practices with high dimensional biomedical data
• dimensionality reduction, visualization and signal feature analysis
• time series analysis
• statistical learning framework (machine learning algorithms) [potential models
  • Generative Models
  • linear regression/mixed effects/multivariate;
  • Bayesian network/autoregressive model
  • Boltzmann machine
  • Discriminative Models
  • logistic regression
  • SVM
  • Neural networks]
• predictive-analytic models: Choosing the Best Model
  • AIC/BIC
• Project wok

Evaluation

• Project based; small groups (O:3)
  • students will make a project plan that includes scope, significance, deliverables and milestones; overall the projects should have several phases
    • phase 1: defining the project, deciding the data, roles, etc.(including a code repo from a predefined list)
    • phase 2: code deliverables:
    • successful contribution to code repo (merged pull request)
    • proper documentation
    • demo/tutorial notebook
    • presentation + live demo
    • update to either CRAN or PYPi
    • phase 3: analysis deliverables:
    • conference report with publication quality figures
    • notebook to reproduce each of the quantitative results
    • presentation
    • results to be deposited in an open access repository
  • evaluation: the progress is evaluated at the end of each phase and the team will be graded for meeting the above deliverables and a clear and concise progress report. the phases are inspired from https://neurodatadesign.github.io/Syllabus/

[SaraMati]

a few notes:

• the project and lectures can go in parallel from the beginning of the term or after the programming section
• the list of statistical models seems long, it doesn't mean that part is the focus of the course
• the focus of the course is not to teach python or R from scratch either. assumption is that either people have a background or will catch up on their own. the main concepts can be reviewed and stressed upon in the first lectures.
• also to teach how to integrate the two languages, e.g . have python for the raw data processing and R for the the statistical methods?

[lwjohnst86]

Amazing! I think any efforts for getting computational training going are needed!

Thoughts:

• The suggested syllabus is a lot that would have to be covered. The key to focus on and remember ... "less is more", especially when it comes to learning. What are 5-7 key skills you want the learners to take away or have at the end of this course?
• If the target learners are "from different backgrounds" but you also suggest "not to teach python or R from scratch either" ... this will be very hard to implement. Many people don't get training in these skills, and then adding the additional layer of coming from different backgrounds, you will quickly end up teaching python and R ... so you need to plan for that.
• I'd recommend a lesson on project setup, workflow management, and "containerizing" their code/data (aka, a repo for the code and maybe one for the data).
• I like the project, team-based approach to the assessment. You could take a lot from the EEB rcourse.
• Since it is a project, team-based assessment, I assume learning to be collaborative is a key skill you want them to develop. If that's the case, the syllabus needs to include some lessons on project management, planning, and understanding and navigating team dynamics. These are not easy skills and they shouldn't be dismissed or diminished.
• "update to either CRAN or PYPi" This might be way outside the scope of a course. We in the rcourse always wanted students to submit to bioRxiv or even figshare, but it doesn't happen. It takes a lot of extra work that learners may not necessarily want to invest in.
• I like the deliverables as part of the assessment. Maybe something a bit lower as "a reproducible, highly polished report" (with submission to get a DOI as an added or bonus thing).
• The focus of the course should be crystal clear: Is it about teaching coding/programming/project setup and management or is it about teaching modelling techniques and dealing with the data? These are two different, highly complex things that take a massive amount of work to teach... It's probably not realistic to teach them both in a single semester course.... especially if the learners come from different backgrounds. Don't underestimate the knowledge level and ability to learn of the students. Again, less is more. Choose one focus area and stick with that. For instance, if focusing on the coding/programming side, you can be general about the statistical model to use so that you can focus on how to write and organize your code and structure your project. Also, which is in more demand and has limited supply? The programming focused course or the modelling/statistical technique focused course?
• To help focus the course down, I would recommend writing a one sentence "elevator pitch" for what a person will learn and take away from the course. The "purpose" and "point" of the course.

If you decide to focus more on the project management/workflow setup, overall code organization and reproducibility side of things, then we could merge some of the things I plan on developing for the "reproducibility and open science toolkit" set of projects.

[QuLogic]

@SaraMati did you see the discuss list posting about https://bioinformatics.ca?

[SaraMati]

@QuLogic I haven't seen any discussions regarding that workshop. where do you refer to?

[SaraMati]

@lwjohnst86 true, the course may be fit for a one year course. this first draft was more for you (UofT council folks) to see if you are up to develop another course based on the r-course.

ans yes, we should teach the languages. This is meant to differentiate the course with the computer science programming courses. I think the approach in teaching programming languages through scientific applications in the carpentaries and the r-course is good.

So, I don't quite understand your comment on separating the coding/programming from modelling/data handling? Isn't the r-course a combination? isn't the approach to teach programming for science, and hence in the lectures we have to talk about some models! I agree that then in the project part we should leave it to students and what is relevant to the projects. [a review of the r-course project can be help].

In general, I agree that the content of the course should be better determined. there are factors that should be determined from the institute side that affects the content:

• if it is one year or one term course
• since the institute is multi-disciplinary, the course can be targeted to (i.e. mandatory for) specific group: like only students from life-sciences or only students from engineering. I will update here upon my meeting with the institute.

Thanks for your input, I assume, the content will form through such discussions between people who sign up for developing the course.

[lwjohnst86]

@SaraMati yes, you are right about the rcourse and the coding vs modelling. I more meant to make sure it is solidly in the area of applying those models in a practical and hands on environment, rather than in the theory and details of the modelling/statistics.

Looking forward to hearing what they say about the idea. As I said, I can contribute material that overlays with what I plan on doing anyway :)