Early warning of neonatal sepsis from vital signs

The problem

Neonatal sepsis develops quickly, and its symptoms only appear once the condition has advanced. Vital signs are already recorded in the incubator every minute, but they're never analyzed: they show on the screen and are lost.

How it works

1. Data collectionPulling the pulse and oxygen time series from the incubator records.
2. CleaningHandling gaps and outliers and unifying the sampling rate.
3. RepresentationExtracting features from time windows: trend, variability, and deviation.
4. AlarmA model outputs a risk score, and defined rules turn it into an actionable alarm.

Project features

• It uses data that already exists, with no new sensors or devices needed.
• The alarm fires hours before clinical diagnosis, and the system's entire value lies in that window.
• Data is split per patient, not per record, to keep the same baby's data from leaking into the test set and producing phantom results.
• Alarm calibration: an alarm every hour makes staff ignore it, so the threshold is part of the design.

Expected challenges

These are the main obstacles to expect during the build, laid out clearly so you know them before you start.

• Patient data leaking between the training and test sets is the biggest risk in this kind of project: the result looks excellent but is wrong.
• Severe data imbalance: cases are very few compared with healthy hours.
• Sensitivity matters more than accuracy: missing a case is far worse than a false alarm.
• Real medical data is hard to access and requires ethical approval.

The consultation gives you a solution for each one.