Blending Knowledge in Deep Recurrent Networks for Adverse Event Prediction at Hospital Discharge

Area of Science:

• Artificial Intelligence
• Health Informatics
• Machine Learning

Background:

• Deep learning models excel at complex data but face challenges with sparse insurance claims data.
• Data sparsity in insurance claims limits deep learning's effectiveness for predicting adverse events like 30-day readmissions.
• Classical machine learning methods often match or exceed deep learning due to handcrafted, domain-specific features.

Purpose of the Study:

• To enhance deep learning capabilities for predicting adverse events using insurance claims data.
• To overcome data sparsity limitations in deep learning for healthcare predictions.
• To develop a hybrid approach combining deep learning with domain knowledge.

Main Methods:

• Introduction of a novel deep learning architecture that integrates patient data representations from a self-attention based recurrent neural network.
• Fusion of deep learning-derived patient data representations with clinically relevant, domain-specific features.
• Extensive experimental validation on a large-scale insurance claims dataset.

Main Results:

• The proposed blended deep learning method demonstrated superior performance compared to standard machine learning approaches.
• The integration of domain knowledge within the deep learning architecture effectively addressed data sparsity issues.
• Accurate prediction of adverse events, including hospital readmissions, was achieved.

Conclusions:

• Blending domain knowledge with deep learning architectures is a viable strategy to improve prediction accuracy for sparse insurance claims data.
• The developed hybrid model offers a promising approach for predicting adverse healthcare events.
• This research highlights the potential of combining artificial intelligence with clinical expertise for enhanced healthcare analytics.