Meltdown Pattern Predictor - Hackathon Backstory
Inspiration
This project is very personal to both of us.
I (Rumaiza) have a brother with severe autism caused by a childhood illness. There are days when he suddenly becomes very stressed or overwhelmed, and as a family we often struggle to understand why. When a meltdown happens, it can feel unpredictable and confusing, even when you try your best to create a calm environment.
Aruzhan has worked with children on the autism spectrum and has seen these challenges firsthand. Through that experience she developed deep empathy for both the children and the caregivers who support them every day.
These experiences made us wonder: what if caregivers could track small daily factors and actually see patterns behind meltdowns? Could simple data help families understand triggers earlier and create more supportive routines?
That question is what inspired us to build Meltdown Pattern Predictor.
What it does
Meltdown Pattern Predictor is a tool that turns guesswork into insight. Caregivers log simple daily factors (how much sleep, how noisy the environment was, whether there was late sugar or screen time, routine changes, late meals) and whether a meltdown happened. Over time, the app shows which conditions tend to show up before meltdowns and which matter most.
It does three main things:
- Daily logging: Quick, structured entries so you don't have to remember everything at the end of a long day.
- Risk prediction: Using your inputs, the app estimates today's meltdown likelihood and highlights which factors are driving the risk. It also lets you play with "What if?" scenarios, like "What if we got 8 hours of sleep instead of 6?" so you can see how small changes might help.
- Pattern analysis: After you've logged a few days, it computes meltdown rates across each factor and surfaces which triggers show up most in your own data.
Everything runs in the browser. No accounts, no servers holding your family's data. We designed it that way because caregiver data is sensitive and private.
How we built it
We kept the front-end simple: vanilla HTML, CSS, and JavaScript, with Chart.js for visualizations. That way the app loads fast, works offline once it's open, and doesn't depend on heavy frameworks.
For the prediction logic, we trained an XGBoost model on synthetic data that we generated based on published autism and developmental research. The synthetic generator models things like sleep, sensory environment, routine changes, and late screens, using coefficients and relationships that show up in the literature. We didn't have access to real caregiver logs, so we built a research-informed dataset that captures plausible patterns. The trained model is exported to a precomputed prediction grid, so the app can run predictions entirely in the browser, no backend needed.
We added an AI help chatbot (OpenAI's GPT-4o-mini) so caregivers can ask questions about meltdowns, autism, or how to use the app. Data is stored locally in the browser.
Challenges we ran into
The biggest challenge was getting useful training data. Real caregivers typically log dozens of entries, not thousands, and we obviously couldn't ask families to share sensitive data for a hackathon. So we had to build a synthetic data generator from scratch, grounding it in peer-reviewed research on sleep, sensory processing, routine changes, and behavior in autistic children. We spent a lot of time reading papers and tuning coefficients so the generated patterns felt realistic and the model's outputs made sense.
We also wrestled with model accuracy. The synthetic generator is built to have bigger variability so the data feels realistic, not overly tidy, but that makes it harder for a model to predict cleanly. We tried four different models, and the highest performance we got was 68% accuracy. It's moderate, but we're hopeful that real caregiver data could change everything.
Another challenge was keeping the app practical. We wanted it to work without accounts or servers, so families could use it immediately, without signing up or trusting us with their data. That meant we had to precompute the prediction grid and handle everything client-side, which limited how dynamic the model could be but made the experience much more accessible.
Accomplishments that we're proud of
We're proud that the model is grounded in real research, not just guesswork. The synthetic data generator and model training pipeline are documented and reproducible, so they can be improved as new studies come out.
We're proud of the privacy-first design. Families can use the app without creating accounts or sending their data anywhere.
We're especially proud of the "What if?" feature. It lets caregivers see how improving one factor, like sleep or reducing noise, would change the predicted risk. That makes the insights actionable, not just informational.
What we learned
We learned how much autism and developmental research exists on sleep, sensory processing, and routine, and how little of it is packaged into tools that families can actually use. Building the synthetic generator forced us to synthesize findings from many papers and think about how factors interact.
We also learned that simple interfaces matter. Caregivers are exhausted. The app had to be quick to fill out and easy to understand, without feeling like homework.
What's next for Meltdown Pattern Predictor
We'd love to support real user data over time, letting the model learn from a family's own logs and personalize predictions. We're also thinking about export options, so caregivers could share anonymized summaries with therapists or doctors.
Longer term, we want to add more factors (like weather, transitions, social demands) and explore whether the app could help with earlier warnings, e.g., nudging "today looks higher risk" so families can plan calmer days when possible.
Most of all, we want to keep it human: a tool that supports caregivers, never judges them, and helps turn confusing days into clearer patterns.
Built With
- and-as-a-family-we-often-struggle-to-understand-why.-when-a-meltdown-happens
- chart.js
- css
- html
- it-can-feel-unpredictable-and-confusing
- javascript
- localstorage
- openai-api
- python
- vercel
- xgboost
Log in or sign up for Devpost to join the conversation.