4th Annual BIDS-TP Hackathon

Schedule

9:00 AM: Breakfast, Team Formation & Project Pitches

9:30 AM: Hacking Begins

12:00 PM: Working Lunch

12:30 PM: Team Presentations

1:00 PM: Wrap Up

Hackathon Projects

Predicting Hospital Readmission from Wearable Data (Aditya Jalin)

Hospital readmissions after bone marrow transplant are costly, stressful, and often preventable. After discharge, patients return home still in a fragile state — their immune systems rebuilding, their bodies adjusting. Some patients deteriorate and need to be readmitted within days or weeks. Could we predict this before it happens?

Wearable devices offer a continuous window into a patient’s physiological state at home. Changes in heart rate, activity level, and sleep may be early indicators that a patient is heading toward a complication. If we can detect these signals early, clinicians could intervene — adjusting medications, scheduling check-ins, or admitting patients before a crisis occurs.

In this project, you will build a machine learning model that uses the first 30 days of wearable data from HCT patients to predict whether they will be readmitted to the hospital. This is a real binary classification problem with clinical stakes

Classic Bioinformatics: Mining Public Genomic Data (Cristina Mitrea)

Option 1 – COVID-19 Transcriptomic Reproducibility: Are COVID-19 gene signatures consistent across independent datasets? Download GEO RNA-seq cohorts, compare differential expression methods, and integrate results across studies.

Option 2 – Prostate Cancer Tumor Microenvironment: Does immune cell composition in prostate tumors shift with clinical risk? Use public single-cell data to map the tumor microenvironment and link cell-type abundance to Gleason score or recurrence.

Both options include a guided analysis pipeline, an original research question you define, and an optional agentic AI component using U-M-provided tools. R or Python. ~3–5 hours.

Curating Bioinformatic Data to be AI Ready (Mathew O'Mera)

Brief task description: In small teams, select a published dataset and prepare it into a HuggingFace dataset with a high-quality dataset card.

There is growing interesting making the diverse and large-scale data generated across the biological sciences available for training machine learning models. While large-scale data repositories such as the SRA/GEO for sequence data, the PDB for structure data, and PubChem and ChEMBL for chemical data provide broad curation of deposited biological data, these databases as well as a wide range of heterogeneous data that is in private databases and supplementary data from papers that are not easy directly include as training or validation data for machine learning models. 

In this hackathon task, each team will select an interesting dataset and make it available as a HuggingFace dataset. Once complete, the data can be loaded into python with 2-lines of code and it is ready and ready to use and for training and testing models. Beyond ease of use, a key requirement of making data AI ready is having a clear description of the providence and scope of the data so that way the models and evaluation tasks that use it can be interpreted correctly. To achieve creating at HuggingFace dataset, we will follow tutorial with best practices to extract/transform/load data, and document in a quick-start guide and dataset card the details of the data.