Stoolum

• 

  Stoolum Logo

• 

  Stoolum UI

• 

  Low Probability Demo

• 

  Medium Probability Demo

• 

  High Probability Demo

Inspiration

• Cirrhosis usually gets caught late, typically once liver damage is severe and treatment options are limited.
• Research on the gut-liver axis shows certain gut bacteria (like Veillonella and Haemophilus) show up more often in people with advanced liver disease
• That led us to a simple question: could a stool sample flag cirrhosis risk before symptoms get serious?
• That was our inspiration for Stoolum: a tool that turns microbiome data researchers already have into a cirrhosis probability score in minutes

What it does

• Estimates cirrhosis probability from stool bacterial abundance data
• Users can upload a MetaPhlAn CSV, enter values manually, or try a demo profile
• Pulls out 10 key bacterial species and shows what it found
• Returns a probability score with a risk level (low/moderate/high)
• Built for research triage/diagnosis

How we built it

• Trained on about 1,989 stool samples with 826 species-level features
• Used ANOVA F-tests to narrow it down to the 10 species most linked to cirrhosis (mostly Veillonella and related oral-gut bacteria)
• A Random Forest classifier learned from those 10 features
• Runs in a Streamlit app that handles messy real-world CSV formats using pandas/numpy
• Shows preview tables so users can trust what's happening under the hood

Challenges

• Microbiome data is sparse; most species show up as zero in most samples, so feature selection had to do a lot of work
• Real CSV exports have long, ugly taxonomy headers, so we built logic to match them to our model's features
• Sliders don't work well for this kind of data; CSV upload became the main path instead

What we're proud of

• A full pipeline, from raw CSV to prediction to readable result
• The tool is built around how researchers actually work, utilizing a MetaPhlAn CSV used in metagenomics

What we learned

• Feature selection matters as much as the model itself when you have way more species than samples
• Good ML tools are workflow tools first: they need to fit into what people already have, not ask for something new
• I learned how to build a Random Forest classifier in Python for the first time.

What's next

• External validation on new data (i.e. testing out the model on different datasets)
• Batch scoring: calculating the scores for multiple samples
• Feature importance explanations so people can see which species drove a given score

Built With

• joblib
• numpy
• pandas
• python
• streamlit