Medi-Care AI: Predicting medications from billing codes via robust recurrent neural networks

Area of Science:

• Health Informatics
• Machine Learning in Healthcare
• Computational Medicine

Background:

• Accurate patient medication lists are crucial but challenging due to data errors and omissions in electronic health records.
• Existing methods struggle with the inherent variability and incompleteness of diagnostic billing code data.

Purpose of the Study:

• To develop a robust deep prediction framework for identifying patient medication classes using sequences of diagnostic billing codes.
• To address data contamination and variability challenges in healthcare records for improved medication prediction.

Main Methods:

• Utilized robust recurrent neural networks (RNNs) for sequence modeling of diagnostic billing codes.
• Implemented an overtime decay mechanism for input billing codes to model temporal patterns.
• Incorporated noise injection into recurrent hidden states, acting as dropout, to enhance model robustness.

Main Results:

• The framework effectively predicts medication therapeutic classes from contaminated diagnostic billing code sequences.
• Demonstrated improved robustness against missing values and multiple errors in healthcare data.
• Validated the effectiveness on real-world healthcare datasets, showing significant improvements in medication order suggestion.

Conclusions:

• The proposed deep prediction framework offers an effective solution for inferring patient medications from noisy billing code data.
• The robust RNN approach enhances the reliability of medication prediction in clinical settings.
• This method has the potential to improve patient safety and treatment efficacy by providing more accurate medication insights.