AMR-Guard

Inspiration

Antibiotic resistance is a WHO top-10 global health threat. It happens when bacteria stop responding to the medicines meant to kill them, and it is driven partly by antibiotics being used when they aren't needed or kept going longer than necessary.

In busy hospital wards, one small step often gets missed: the antibiotic review. Two or three days after starting treatment, the doctor should stop and ask, "should we continue, change or stop?" Lab results may have come back by then, but with dozens of patients, the review can slip. Small hospitals also rarely know which antibiotics still work in their own area.

We wanted to build something that helps doctors at exactly this moment, without ever making the medical decision for them.

What it does

AMR-Guard is an assistant for hospital doctors and infection-control teams. It never diagnoses, never prescribes and never suggests a drug or dose. Every decision stays with the doctor. It has three layers:

1. Smart review list. Each day, it shows the ward doctor a short ranked list of patients whose antibiotic most needs a review, with plain reasons such as "day 3 of treatment, culture result is back, patient stable for 48 hours."

2. Local resistance chart. It builds a simple chart of which antibiotics are still working in that hospital, using the hospital's own lab results. It also shows how confident the estimate is: "works in about 70% of local cases, but based on only 12 samples, so low confidence."

3. Early outbreak alert. If the same resistant bacterium shows up in several patients in one ward within a few days, it alerts the infection-control team so they can act early.

How we plan to build it

• Review ranking: simple, transparent rules first (days on antibiotic, culture status, patient stability), with every flag explained. We would add machine learning only once real data exists.
• Honest resistance estimates: the tool starts from wider data and shifts toward the hospital's own results as more lab tests arrive. It always shows how confident it is, so small hospitals aren't misled by a few samples.
• Outbreak detection: a simple statistical check that compares recent case counts in a ward with its normal level and raises an alert when they jump.
• Privacy: patient data stays inside the hospital, and reports use anonymised numbers.

We would start with public and synthetic datasets, then test in one or two hospital wards with clinician feedback.

Challenges we expect

• Messy hospital data: we start with a small set of fields (drug, start date, culture result) instead of everything.
• Alert fatigue: too many reminders get ignored, so we limit the list to the few most important patients per day.
• Small sample sizes:we show uncertainty openly and avoid pretending to know more than we do.
• Safety and trust: the tool only informs. Each flag shows why it appeared, and the doctor stays in control.

What we learned

• Some of the most useful health technology doesn't predict or diagnose anything. It makes sure the right information reaches the right person at the right time.
• Being honest about uncertainty is a feature, not a weakness.
• A problem as large as antibiotic resistance can be tackled through many small, well-timed actions in everyday hospital care.

What's next

1. Build a working prototype on public and synthetic data.
2. Get feedback from doctors and infection-control staff.
3. Pilot in a hospital ward and measure whether antibiotic reviews happen more often and on time.
4. Test how few lab samples a hospital needs before its local resistance estimates become reliable, which could become a research paper.

Built With

• docker
• fastapi
• pandas
• plotly
• postgresql
• python
• react
• scikit-learn
• scipy
• sql
• statsmodels
• streamlit
• synthetic-data
• xgboost