Predictive Analytics for Care and Management of Patients With Acute Diseases: Deep Learning-Based Method to Predict...

Area of Science:

• Computational medicine
• Machine learning in healthcare
• Predictive analytics for acute diseases

Background:

• Acute diseases present rapid complications impacting patient outcomes.
• Predictive analytics can aid timely diagnosis and treatment of acute complications.
• Challenges include limited early-stage data, temporal heterogeneity, and imbalanced outcome distributions.

Purpose of the Study:

• To propose a novel deep learning method for predicting acute disease complication phenotypes.
• To address data insufficiency, temporal heterogeneity, and imbalanced outcomes in patient data.
• To enhance the accuracy of predicting crucial complication phenotypes in acute illnesses.

Main Methods:

• Developed a deep learning model using recurrent neural network-based sequence embedding for disease progression.
• Incorporated a latent regulator to manage data insufficiency and unobserved underlying mechanisms.
• Employed cost-sensitive learning to handle imbalanced outcome distributions and improve prediction performance.

Main Results:

• The method achieved an AUC of 0.82 for acute hepatic encephalopathy prediction, outperforming benchmarks by up to 64%.
• For hepatorenal syndrome prediction, the AUC was 0.64, surpassing temporal MMCBR by 29%.
• Demonstrated significant improvements in recall, F-measure, and AUC compared to benchmark techniques.

Conclusions:

• The proposed method effectively learns short-term temporal representations for predicting complication phenotypes.
• It offers superior predictive utility over existing data-driven techniques for acute diseases.
• The approach is generalizable to various acute conditions with data limitations and heterogeneity.