ICU Life Resource Optimiser
A predictive system that analyses ICU patient data within the first 48 hours to forecast mortality risk and length of stay, enabling hospitals to prioritise resources and maximise patient survival outcomes.
Context
Critical care units face constant challenges in balancing limited ICU beds, staff attention, and life-saving equipment. Patients admitted often require urgent prioritisation, but manual evaluation can be inconsistent and resource-intensive.
The ICU Life Optimiser leverages patient data (static and temporal within 48 hours) to generate predictions of in-hospital mortality and length of stay, giving healthcare teams actionable insights to make faster, more accurate allocation decisions.
Problem
ICU congestion, unpredictable patient outcomes, and limited staff resources make it difficult for hospitals to allocate care efficiently. Physicians must often rely on subjective judgment or incomplete information in the first hours of admission, leading to risks of under-monitoring critical patients or inefficient use of ICU beds.
Without a reliable system to evaluate severity early and predict outcomes, hospitals struggle to balance life-saving priorities with operational constraints.
Given the first 48 hours of ICU patients' data, predict per patient their:
In-hospital mortality (1: will die, 0: will not die)
Length of stay (days in hospital)
So as to evaluate the severity of patients and priortise them for optimal resources allocation to maximise monitoring and support to those who need them more.
Static Data that are collected at the time the patient is admitted to the ICU:
RecordID (a unique integer for each ICU stay)
Age (years)
Gender (0: female, or 1: male)
Height (cm)
ICUType (1: Coronary Care Unit, 2: Cardiac Surgery Recovery Unit, 3: Medical ICU, or 4: Surgical ICU)
Weight (kg)*.
We designed a dual-model predictive system combining regression and classification to assist ICU resource planning:
Regression model → predicts length of ICU stay (in days).
Classification model → predicts in-hospital mortality (Will Die / Will Not Die).
Deployment Workflow
Data extraction → Static (age, gender, height, weight, ICU type) and temporal variables are taken from the first 48 hours of ICU stay.
Preprocessing → Missing values (–1) are imputed with the most frequent values across folds; data normalised and dimensionality reduced.
Dataset preparation → Data split into features (X) and labels (Y: mortality, length of stay), then partitioned into training and test sets with 4-fold cross-validation.
Model training & evaluation →
Regression: evaluated on RMSE and adjusted R².
Classification: evaluated on ROC-AUC, accuracy, precision, recall.
Deployment cycle → Every 48 hours, new patient data is processed and predictions are generated:
Regression model forecasts expected stay (helps schedule ICU capacity).
Classifier flags mortality risk (prioritises monitoring & interventions).
Decision support → Hospital managers allocate beds, staff, and resources dynamically, with higher support for patients flagged at risk.
This workflow ensures the system is not just an algorithm, but a practical tool integrated into ICU decision-making.
Models were validated with 4-fold cross-validation to test generalisability.
Regression model performance
In our evaluation of the regression models:
Best model: Multilinear Regression (Design Matrix 3).
RMSE = 19.795 (test), Adj. R² = 0.108.
Insight: modest predictive power; demonstrates feasibility but limited accuracy due to data constraints.
Classifier model performance
In our evaluation of classifier model:
Best model: Logistic Regression pipeline (Scaler + SelectKBest + PCA).
ROC-AUC = 0.817 (train), 0.613 (test).
Accuracy = 87.45%, Precision = 61.67%, Recall = 25.16%.
Insight: reliable accuracy, but recall reveals under-detection of actual deaths (a critical limitation).
Built a functional prototype pipeline for both mortality and length-of-stay predictions.
Validated feasibility of leveraging patient data to inform ICU resourcing decisions.
Demonstrated potential for integration into hospital systems, while also exposing key challenges such as model generalisability, recall gaps, and data limitations.
Length of stay is harder to model → influenced by complex, unstructured factors beyond initial data.
Operational integration matters as much as modelling → hospitals need clear workflows to act on model predictions.
Future opportunity → expanding data inputs and piloting in real hospital settings could transform ICU resource allocation.