Predicting Abnormalities in Laboratory Values of Patients in the Intensive Care Unit Using Different Deep Learning...

Area of Science:

• Medical Informatics
• Artificial Intelligence in Healthcare
• Clinical Decision Support

Background:

• Rapid growth in medical knowledge and health data, including electronic health records (EHRs).
• Increasing complexity of evidence-based medicine and data overload for healthcare workers.
• Risk of overlooking critical patient data and trends due to irrelevant information in EHRs.

Purpose of the Study:

• To analyze ICU laboratory results and classify potential near-future abnormalities.
• To support clinical decision-making by highlighting critical lab values and future testing needs.
• To improve efficiency and allow clinicians more time for direct patient care.

Main Methods:

• Extracted 25 lab values for mechanically ventilated ICU patients from MIMIC-III and eICU datasets.
• Applied time-windowed sampling, holding, and support vector machines for data imputation.
• Utilized Tukey range for anomaly detection and deletion, followed by training deep learning and gradient boosting models.

Main Results:

• Models achieved at least 80% accuracy on multi-label classification of abnormal lab values.
• A multiple convolutional neural network model demonstrated superior performance, exceeding 89% accuracy on both datasets.
• The integrated preprocessing pipeline and machine learning models proved effective.

Conclusions:

• Machine learning and deep neural networks can effectively predict near-future abnormalities in laboratory values.
• The developed system, validated on extensive datasets, shows potential for real-world EHR application.
• The model can enhance patient diagnosis and treatment by improving the interpretation of clinical data.