Final Project

The final project is an opportunity to follow your curiosity, go deep in a particular healthcare domain and dataset, and build something interesting. You are responsible for coming up with a project idea and delivering a 1-page proposal to the course staff by Milestone 1 (see below), and iterating on the proposal until you receive approval from the instructor. You will then implement the project and present it to the course staff and a broader CMU audience in a public poster session, similar to what you would experience at an academic conference.

It is NOT required that you develop something completely novel and publishable. A perfectly acceptable, full-credit final project would be to replicate (perhaps with your own twists) an existing result from the literature. Before submitting your proposal to the course staff, be sure to ask yourself these questions:

1. Does this project address a meaningful problem in healthcare? (The first third of the course will help you make an informed assessment.)
2. Does this project rely on machine learning (supervised, unsupervised, or RL)? Healthcare analytics projects are important and interesting, but this is an ML class.
3. Does this project leverage a healthcare dataset?
4. Is this project non-trivial? (E.g. downloading and running the code and data from a public github repo associated with a published paper would not pass this criterion)

While we don't want to be overly-prescriptive about the project writeup, you will want to include at least some of the typical elements of most papers published in the last five years in the area of machine learning in healthcare, such as:

• EDA (exploratory data analysis) and a "Table 1" describing your data in detail
• Pointer to source code (for reproducibility)
• Why you selected the model form(s) you did
• Some details on model building: train/test split, approach to smoothing, number of iterations, etc.
• Where and how your work fits relative to the academic literature
• Shortcomings of your work and recommendations for future work

Sample ideas

Here’s some example ideas for a project - to get you calibrated on the scope and complexity we expect.

• Pneumonia Detection Challenge: Participate retrospectively in the 2018 RSNA Pneumonia challenge.
• Detecting abnormal heartbeat: Participate retrospectively in the 2016 Physionet challenge, to classify recordings of heartbeats. Physionet runs a cardiology-related challenge every year and the previous year challenges, along with data, are available here.
• Automated Medical Coding: Using MIMIC or another dataset, build a model that assigns ICD-9/ICD-10 codes based on the clinical information (including notes) generated during a visit. See here, among many other sources.
• Question answering (Q/A) on EMRs: The typical EMR patient record contains a vast amount of information, especially for older patients with chronic diseases. A high-quality Q/A system can allow physicians to to quickly retrieve specific information from a patient chart by asking natural language questions, instead of manually searching through pages and pages of information. Use a base model like ClinicalBERT and attach a Q/A head on top of this model, which you will probably want to train using an annotated clinical question-answering dataset (e.g. emrQA).
• Patient embedding: Loosely following the work described in the DeepPatient paper, build an embedding model for patients and demonstrate its performance on some healthcare application.
• Quality of care assessment: Use CMS data to build a model that predicts the quality of care provided by healthcare providers and institutions. (You will need to define ‘quality’). You may wish to look at datasets containing, for example, hospital readmissions, mortality rates, and patient satisfaction data. What features are most associated with high-quality care?
• Multimodal AI: There are more and more public multimodal healthcare datasets becoming available, consisting of images and associated metadata. One example is the ISIC dataset, which contains images of skin lesions along with corresponding clinical information. Use a pre-trained convolutional neural network (CNN) model, such as ResNet, to extract visual features from the dermatoscopic images. In parallel, use a pre-trained language model, such as BERT or BioBERT, to extract textual features from the clinical notes. Explore different techniques for fusing the visual and textual features.
• For other ideas for a project, have a look at this survey paper.

Some popular, public healthcare datasets you may wish to consider:

1. MIMIC
2. The Cancer Imaging Archive
3. The National Cancer Institute ‘SEER’ data
4. Human Connectome Project (HCP)
5. CDC data
6. CMS data
7. Human Mortality Database
8. Chexpert chest x-ray data
9. Kaggle also has many healthcare datasets and so does this.

Some of these datasets may require paperwork/approval to get access, so plan ahead.

Milestones

The Final Project will have three milestones during the semester:

• Milestone 1 (Oct 10): Submit 1-page project proposal, which describes the problem you are addressing, a brief outline of your approach, and the data you intend to use.
• Milestone 2 (Nov 7): Submit 1-page progress report, describing work completed and pending.
• Milestone 3 (Dec 3): Poster session / live demo. Submit your poster as a pdf within Canvas before the poster session.