Back to mahak's profile
18views

Home Ward

A home “hospital ward” that senses, predicts and alerts before a crisis, even when the internet or power fails.

mahak fahad
  • Video of it working
  • Problem, approach and result explained
  • Public source code
  • Commits by the candidate· 11 of 11 commits
  • Built during the challenge
  • Answered the full brief· 6 of 6
  • Expert reviewed
Home Ward

10 h

earlier warning than standard scoring (simulated test patients)

0.115

false alerts per stable patient-day, vs 0.265 for the standard score

166

automated tests (Python and real-browser), run on every push

Walkthrough

Overview

The problem

After a patient goes home from hospital, the monitoring usually stops. A family member watches for problems with a thermometer, and deterioration is noticed late. Cheap home monitors raise so many false alarms that families start ignoring them, and when Wi-Fi or power fails, they go silent. We set out to build a home that works like a small hospital ward: it notices slow changes early, explains why, and makes sure a real person responds, even when the internet or power fails.

The approach

Sensing and hub: a simulated bedside hub per home (homeward/edge.py) streams vitals, pill box, bed and motion sensors, oxygen-machine data and power status. It stores readings first and replays them after an outage. Risk: every reading is scored with NEWS2, the score hospitals use, including the COPD scale for oxygen patients. Each patient has a personal baseline. An explainable model (homeward/model.py) forecasts the chance of deterioration in the next 24 hours, and each factor’s contribution is shown. Languages: caregiver notes are understood in English, Hindi, Hinglish, Tamil and Tanglish, including negation such as “bukhar nahi hai” (homeward/notes.py). Alerts and privacy: alerts are confirmed on two readings and escalate on a timer. Family members see only their own patient. The hardest part: making alerts trustworthy. Sensor glitches, one-off spikes and duplicate data all had to be filtered out so that alerts mean something. We handled this with a signal-quality check, two-reading confirmation and deduplication.

The result

It runs with one command (python run.py) and needs no hardware. The live demo is at https://homeward-itmp.onrender.com. Against held-out simulated patients, the early warning came a median of about 10 hours before NEWS2 alone, with fewer false alarms (0.115 vs 0.265 per stable patient-day). Sudden emergencies can’t be predicted, and the system says so. Testing covers 144 Python tests and 22 real-browser checks, all run on every push through GitHub Actions.

Reflection

Some things could be better. The model is trained on simulated patients, so it needs real data and clinical validation before anyone relies on it. Free hosting sleeps when idle, so the first visit is slow. The demo passwords are visible on the sign-in page, which is fine for a demo but not for real use. Next, we’d add real sensor hardware, connect to hospital records and collect feedback from real caregivers. The hardest lesson was that reliability and explanation matter more than adding features.

AI tools used

Claude

Drafting and reviewing code, writing tests and documentation, and generating the voice-over. The team designed the clinical logic and reviewed every change

Links & files

Artifacts

4

Gallery

5