Quick Questions for April 23, 2020

JavaScript isn't enabled in your browser, so this file can't be opened. Enable and reload.

These questions are answered within the lecture video, and we hope they will help you stay on track watching lecture and make sure you're processing the content! Please complete by Thursday at midnight, Eastern time.

Email *

MIT Kerberos (part in email before @mit.edu) *

Disease subtyping refers to... *

1 point

Breaking down a broad disease into tighter, ideally biologically defined subgroups based on underlying physiology and/or clusters within the disease trajectories

Narrowing down within a research project the defining features of patients of interest for a study

Examining how disease burden disproportionately affects different groups of people

Analyzing how a disease has been studied historically to identify different new possible research directions

What are some reasons why you may want to tie weights together (i.e., use multi-task learning) when performing prognostication or disease progression modeling? Check all that apply.

1 point

By aggregating over timepoints (tasks) we account for data sparsity & imbalance.

Multi-task learning over N tasks lets you learn an N-times as powerful model at no extra cost.

Multi-task learning allows us to more efficiently approximate the ultimately better markov model approach without the extra cost.

Multi-task learning can help us learn more robustly in the presence of censoring.

Multi-task learning can help us regularize over label noise.

When trying to perform multi-task learning for disease progression modeling over an ordered sequences of timepoints t_1, ..., t_n (e.g., predict disease state at 6 months, 1 year, 18 months, etc.) with linear models, you might consider (select all that apply) *

1 point

Jointly optimize for all linear models simultaneously, keeping the weights unconstrained for all.

Penalizes the discrepancies between the weights of each pair of consecutive models (e.g., ensure that the weights for the 12 month predictor are close to those of 6 month and 18 month models).

Penalize the discrepancies between any weight vector and a dummy, sentinel weight vector w_* which you add to the problem as a regularizing center.

Penalize the discrepancies between any weight vector and the weight vector corresponding to t_1, as this will force all weight vectors together anyways.

Required

What are some reasons one might want to perform unsupervised disease progression modelling?

1 point

Clustering patients into meaningful splits can help us identify cohorts for clinical trials.

We may not have enough labeled data for the task at hand but still want to perform some form of disease progression modelling.

We'd like to better understand the disease mechanistically, and can use raw clusters to motivate scientific research.

We'd like to build a strong dimensionality reduction technique to enable downstream analysis.

Submit

Clear form

Never submit passwords through Google Forms.

This content is neither created nor endorsed by Google. Report Abuse - Terms of Service - Privacy Policy

Forms