Sepsis Early Warning From Continuous Vitals

Standard bedside sepsis scores can identify patients late. The question was whether continuously collected bedside signals could provide an earlier, broadly available warning.

I analyzed 1,552,210 hourly clinical readings across 43 measures. Because 27 measures were more than 90% incomplete, I removed them rather than impute unreliable data. I then compared 6 model types and narrowed the system to 8 continuously available inputs.

1,552,210 hourly readings processed

43 clinical measures evaluated

8 inputs selected for an every-bed deployment path

The result is intended to prioritize patients for review, not act as an autonomous alarm. It was built on one public, de-identified ICU dataset and needs external validation.