Long Short-Term Memory Recurrent Neural Networks for Multiple Diseases Risk Prediction by Leveraging Longitudinal...

Area of Science:

• Medical Informatics
• Computational Health

Background:

• Chronic diseases pose a significant societal burden due to delayed identification.
• Longitudinal patient data offers potential for proactive disease risk assessment.

Purpose of the Study:

• To develop and validate a multiple disease risk prediction method using longitudinal electronic health records.
• To assess the performance of the prediction model at different levels of diagnostic code aggregation.

Main Methods:

• Utilized recurrent neural networks with long-short time memory units for risk prediction.
• Aggregated International Classification of Diseases (ICD) codes to different granularities (3-digit and 4-digit) for analysis.
• Validated the model on two independent hospital datasets comprising over 20,000 patient records.

Main Results:

• The recurrent neural network model demonstrated high accuracy in predicting future disease risks.
• Achieved exact-match scores of up to 98.90% (3-digit ICD) and 96.83% (4-digit ICD) across datasets.
• Identified significant variations in patient characteristics necessitating tailored prediction approaches.

Conclusions:

• The developed multiple disease risk prediction method is effective for proactive patient healthcare management.
• The model shows strong applicability for integration into hospital information systems.
• Accurate disease risk prediction can help mitigate the societal burden of chronic diseases.