An artificial neural network approach to diagnose and predict liver dysfunction and failure in the critical care...

Detecting liver dysfunction/failure in the intensive care unit poses a challenge as individuals afflicted with these conditions often appear symptom-free, thereby complicating early diagnoses and contributing to unfavorable patient outcomes. The objective of this endeavor was to improve the chances of early diagnosis of liver dysfunction/failure by creating a predictive model for the critical care setting. This model has been designed to produce an index that reflects the probability of severe liver dysfunction/failure for patients in intensive care units, utilizing machine learning techniques.

This effort used comprehensive open-access patient databases to build and validate machine learning-based models for predicting the likelihood of severe liver dysfunction/failure. Two artificial neural network model architectures that derived a novel 0-100 Liver Failure Risk Index were developed and validated using the comprehensive patient databases. Data used to train and develop the models included clinical (patient vital signs) and laboratory results related to liver function which included liver function test results. The performance of the developed models was compared in terms of sensitivity, specificity, and the mean lead time to diagnosis.

The best model performance demonstrated an 83.3 % sensitivity and a specificity of 77.5 % in diagnosing severe liver dysfunction/failure. This model accurately identified these patients a median of 17.5 hours before their clinical diagnosis, as documented in their electronic health records. The predictive diagnostic capability of the developed models is crucial to the intensive care unit setting, where treatment and preventative interventions can be made to avoid severe liver dysfunction/failure.

Our machine learning approach facilitates early and timely intervention in the hepatic function of critically ill patients by their healthcare providers to prevent or minimize associated morbidity and mortality. HIPPOKRATIA 2024, 28 (1):1-10.