Condition-Aware ICU Decision Support System

The clinical picture

An intensive care unit is a noisy place. Monitors alarm constantly, and a large share of those alarms are not clinically actionable.

The consequence is well documented and has a name: alarm fatigue. When staff are interrupted often enough by alerts that turn out to be nothing, the alerts stop carrying urgency, and the one that mattered gets treated like the ones that did not.

Early-warning systems that flag deteriorating patients are meant to help. They can also make the problem worse, because a system that cries wolf adds to the noise it was supposed to cut through.

Why it's hard

Most early-warning models treat every patient the same way. They learn a general pattern of deterioration and apply one threshold across the unit.

But a vital sign reading that is alarming in one patient is unremarkable in another, depending on why they are in the ICU, what they are being treated with, and what their baseline looks like. A model that ignores that context will keep firing on patients whose numbers are abnormal for the population but normal for them.

That is a false-positive problem, and in an ICU false positives have a cost measured in attention rather than in treatment.

The idea

This system conditions the early-warning signal on the patient's clinical situation rather than scoring everyone against one population-level pattern.

The aim is a smaller number of alerts that are more often worth acting on, which is a different objective from maximizing raw sensitivity.

How it works

The framework is built on MIMIC-IV intensive care time-series, combining vital signs, laboratory abnormalities, diagnosis embeddings, and treatment exposures so the model has access to the context that makes a reading meaningful.

It is structured as a modular pipeline: data processing, feature engineering, modeling, evaluation, and deployment as independent components. The serving layer is a Dockerized API with WebSocket streaming, so inference behaves like a bedside monitor rather than a batch job.

Reproducibility was a design constraint. The project is split into components that can be rerun independently, and no raw patient data is included in the repository, in line with PhysioNet data governance requirements.

What happened

The pipeline runs end to end, from raw ICU time-series through to real-time streaming inference.

This is ongoing work and the evaluation is not finished. Rather than quote provisional numbers here, the honest statement is that the infrastructure and the conditioning approach are built and the clinical evaluation is still in progress.