Prediction of ICU Patient Deterioration: Application of Machine Learning in Early Clinical Intervention

This open-source project was released by Aishwarya-1367 on GitHub on June 8, 2026 (link: https://github.com/Aishwarya-1367/ICU-Patient-Deterioration-Prediction). Its core goal is to use Random Forest, CatBoost, and Gradient Boosting algorithms to identify high-risk patients with deteriorating conditions from ICU patients' clinical and physiological data, providing data support for early clinical intervention to improve patient outcomes.

The Intensive Care Unit (ICU) is a medical environment in hospitals that requires close monitoring. Patients' conditions can deteriorate rapidly in a short time, so timely detection of early warning signals is crucial for saving lives. Traditional clinical monitoring relies on medical staff's experience and regular rounds, but human resource limitations may lead to subtle changes being overlooked. This project explores the use of machine learning technology to automatically identify high-risk patients, supporting early intervention and improving outcomes.

The project uses a public medical dataset from Kaggle, which includes multiple physiological indicators and clinical records of ICU patients. Data preprocessing includes cleaning, missing value handling, feature selection, and engineering. To address the class imbalance issue in medical data where deteriorating cases are far fewer than stable cases, a specialized sampling strategy is used. Feature engineering extracts key indicators such as vital sign trends, laboratory results, and basic health status for model training.

The project compares three machine learning algorithms:

1. Random Forest: An ensemble learning method that builds multiple decision trees for comprehensive prediction, stable in handling high-dimensional data and not prone to overfitting;
2. CatBoost: A gradient boosting algorithm developed by Yandex, which excels at handling categorical features (such as diagnostic codes, treatment plans);
3. Gradient Boosting Classifier: Sequentially builds weak learners to correct previous errors, performing excellently in structured data competitions. Model performance evaluation metrics include accuracy, precision, recall, F1 score, and the area under the ROC-AUC curve.

The model achieved an overall accuracy of 93% on the test set and an ROC-AUC of 0.895, but detailed analysis reveals typical challenges:

Metric	Stable Class (0)	Deteriorating Class (1)
Precision	0.94	0.69
Recall	0.98	0.36
F1 Score	0.96	0.47
The model performs excellently in identifying stable patients (recall of 0.98), but its recall for detecting deteriorating patients is only 0.36—over 60% of deteriorating cases are missed. This is because deteriorating cases are rare, and the model tends to predict most patients as stable to achieve high overall accuracy, sacrificing recall for the minority class.

The project has the following limitations:

1. Class imbalance in the dataset affects the ability to identify deteriorating cases; future attempts can include sampling techniques like SMOTE or cost-sensitive learning;
2. The model has not been deployed and tested in a real-time clinical environment; the gap between laboratory performance and actual deployment needs further verification;
3. The model's generalization ability may vary due to differences in hospitals, patient populations, or regions, requiring continuous research.

Despite its limitations, the project demonstrates the application potential of machine learning in the medical field:

• Developers: Provides a reference for the complete workflow of medical machine learning (data acquisition, preprocessing, feature engineering, model training and evaluation), especially the methodology for handling class imbalance and comparing multiple models is of reference value;
• Clinical Workers: AI can be part of an early warning system to help prioritize high-risk patients. Although it cannot replace clinical judgment, it assists in decision-making; This project represents an important direction for medical AI. With data accumulation and algorithmic progress, it is expected to become part of standard ICU care, providing additional protection for patient safety.