JMIR Publications

ABSTRACT

Background:

Cirrhosis is a chronic liver disease, manifested by fibrosis and regenerative nodules. It can be classified as compensated or decompensated clinically. Accurate predictions of patients’ transition from compensated to decompensated cirrhosis are important but difficult, due to the multifactorial nature of disease progression, comorbid conditions, and patient heterogeneity.

Objective:

To effectively identify at-risk cirrhotic patients and facilitate timely interventions across chronic diseases and patient populations, this study develops a novel, deep learning–based method to predict patients’ transition from compensated to decompensated cirrhosis. The proposed method provides superior, predictive utilities and can help physicians derive actionable clinical insights by modeling comorbidity relationships and symptom interdependencies that influence the progression of cirrhosis among patients.

Methods:

The proposed method is guided by clinical knowledge, considers patient demographics and medications, and leverages a graph structure to capture important comorbidity interplays. This method is also augmented with post-hoc interpretability to help clinicians understand its predictions for clinical decision-making and patient management. Using clinical data pertaining to 29,915 cirrhotic patients, we evaluate the proposed method in comparison with 12 prevalent methods.

Results:

The proposed method attains an average area under the curve (AUC) of .87, F-measure of .681, recall of .741, and precision of .662. It achieves a 6.1% improvement in AUC and a 14.2% improvement in F-measure over the best-performing benchmark method. Supported by post-hoc interpretation analyses, this method can provide physicians with interpretable visualizations of the underlying comorbidity and symptom dynamics to better support their decision-making and adoption.

Conclusions:

The empirical results underscore the value of demographic similarity, comorbidity graphs, and extended loss functions for identifying cirrhotic patients at higher risk of decompensation by capturing finer-grained information from patient features that tend to be overlooked by most existing methods. The proposed method not only supports physicians’ clinical decision-making and patient management, but also can enhance healthcare organizations’ resource allocation efficiency and cost containment.