Mammoly

BioHacks Tracks:

Mammoly was built with purpose across all three hackathon tracks. On the Bioethics and Justice track, every AI decision is fully explainable, every known bias is named, and a dedicated accountability tab keeps the tool honest and transparent. On the Cancer Applications in Genomics track, our model was trained on 1.5 million real clinical records to understand how personal risk factors — like breast density, family history, and hormonal history — shape each patient's unique cancer risk. On the AI Applications in Biomedical Research track, Mammoly goes beyond prediction by auditing its own fairness, simulating the effects of health policies, and connecting underserved patients to free screening resources.

Inspiration:

Our inspiration comes from a close family member who had breast cancer. During that time, we researched and learned how critical early detection is, as the survival rate is near 100% for Stage I but drops dramatically for Stage IV. We also found inconsistent screening recommendations across authorities and physicians, along with clear disparities in mammogram rates by race, insurance status, education, and immigration background. Our family member survived, but that experience showed us firsthand how much uncertainty, fear, and delay can cost a family even when the outcome is good. This project integrates AI into clinical decision support to personalize screening intervals, audit bias, and optimize equitable policy planning. It is not designed to replace physicians, but to ethically support their judgment while addressing the systemic inequities that cause too many people to receive a diagnosis far later than they should have.

What it does

Mammoly is a breast cancer risk prediction and screening equity platform built on the Breast Cancer Surveillance Consortium (BCSC) dataset of 1.5 million clinical observations. It is designed to live inside a doctor's portal as a clinical decision support tool. It has three core functions: a Patient Risk Recommender that takes 10 clinical inputs (age, breast density, family history, BMI, HRT use, menopausal status, and more) and returns a calibrated risk probability, a screening interval recommendation, and a feature by feature explanation of what's driving the score. A Bias Audit Dashboard that automatically slices model performance across age groups, racial/ethnic groups, and breast density categories, surfacing disparity alerts when gaps exceed clinical thresholds. For patients who face financial or access barriers, Mammoly also surfaces free and low cost screening resources so no one leaves without a next step. Additionally, an Equity Screening Optimizer where policymakers can simulate the population level impact of changing screening guidelines, like lowering the start age from 50 to 35, and see exactly how many additional cancers would be detected and how many additional false positives would occur, broken down by racial subgroup.

How we built it

We started with the three BCSC summary datasets, concatenated them, and expanded the count weighted records into 6.7 million individual observations. We cleaned the data, handled missing values, and trained a logistic regression model that not only predicts risk but can explain its reasoning in plain terms. The frontend is Streamlit with fully custom CSS, pulling the color palette from prior research we have conducted in regards to Breast Cancer screening. We also created a custom logo, and used React Bits to add an active background to make it appealing for users. We also built a bias testing engine that automatically checks model fairness across demographic groups, and a simulation tool that estimates real world policy impacts.

Challenges we ran into

The dataset was heavily imbalanced since only about 7% of patients had a cancer history, which meant a lazy model could look accurate while being completely useless. We had to use special weighting techniques to fix that. The full dataset also expanded to nearly 7 million rows which was too large to train on quickly, so we developed a sampling approach that kept the results accurate while being fast enough for a live demo. The trickiest part was figuring out what fairness even means in a medical context, because simply treating all groups the same statistically is actually wrong when the underlying health differences are real.

Accomplishments that we're proud of

We built a working, explainable AI model on real clinical data in a single weekend and hit a strong accuracy score of 0.864 AUC. More importantly the bias audit revealed something genuinely meaningful: the model performs noticeably worse for Native American patients than for white patients, which mirrors the real world gap where Native American communities have the lowest mammogram rates of any group in our research data. We're proud that Mammoly isn't just a technical demo: it is backed by real research, insights from an expert interview with a breast health nurse navigator, and real clinical data.

What we learned

We learned that fairness in healthcare is a clinical problem before it's a statistical one. A model with low sensitivity in an underserved group doesn't just have worse metrics, it means real patients with cancer go undetected. We also learned that naming limitations clearly is more credible than hiding them. The BCSC data predicts prior cancer history rather than prospective risk, it lacks BRCA genetic data, and it over represents certain registry populations. Being explicit about all of that and showing we designed around it made the project stronger. And we learned that making equity quantifiable, showing a specific number of additional cancers detected per subgroup under a given policy, is what turns a research finding into something actionable.

What's next for Mammoly

We want to integrate real time patient data so that risk scores update continuously as new clinical information comes in rather than relying on a static snapshot. We also plan to add more predictive features like genetic risk factors such as BRCA status, lifestyle data, imaging history, and environmental factors that can meaningfully improve how accurately we identify who is at risk. On the fairness side we want to connect Mammoly to prospective outcome data so predictions are based on future risk rather than past history. Long term we'd love to see Mammoly fully embedded in doctor portals across different health systems, with the bias auditing framework open sourced so other healthcare AI teams can build on it too. The core mission stays the same: early detection saves lives, and Mammoly is here to make sure it saves them equally.