Enhancing Patient Outcome Prediction Through Deep Learning With Sequential Diagnosis Codes From Structured Electronic...

Area of Science:

• Artificial Intelligence in Healthcare
• Biomedical Informatics
• Machine Learning for Clinical Prediction

Background:

• Structured electronic health records (EHRs) generate vast amounts of patient data, including sequential diagnosis codes.
• Temporal medical history from diagnosis codes is valuable for predicting patient outcomes.
• The integration of sequential diagnostic data into deep learning (DL) models remains underexplored.

Purpose of the Study:

• To systematically review the use of sequential diagnostic data in DL models.
• To understand data integration methods, the impact of sample size on performance, and model generalizability.
• To identify trends in DL architectures and prediction tasks for clinical outcomes.

Main Methods:

• Systematic literature search across PubMed, Embase, IEEE Xplore, and Web of Science up to May 15, 2023.
• Inclusion of studies using DL algorithms trained on sequential diagnosis codes for outcome prediction.
• Evaluation of DL techniques, datasets, prediction tasks, performance, generalizability, explainability, and risk of bias using PRISMA and PROBAST tools.

Main Results:

• 84 studies met eligibility criteria, with a yearly increase in publications.
• Recurrent neural networks (56%) and transformers (26%) were dominant DL architectures.
• Larger training sample sizes correlated positively with model performance (p=.02), but 70% of studies had a high risk of bias.

Conclusions:

• Deep learning holds significant potential for predicting patient outcomes using sequential medical codes.
• Improved predictive performance is associated with multi-feature use, time interval integration, and larger sample sizes.
• Few studies addressed generalizability (8%) or explainability (45%), highlighting areas for future research.