Identifying Disease of Interest With Deep Learning Using Diagnosis Code.

Area of Science:

• Artificial Intelligence
• Machine Learning
• Deep Learning

Background:

• Autoencoders (AE) are deep learning models that reconstruct input data.
• This study explores using AEs for disease prediction based on diagnostic codes.

Purpose of the Study:

• To develop and evaluate a novel supervised autoencoder model for predicting disease co-existence.
• To assess the model's performance using only diagnostic codes.

Main Methods:

• Trained and tested two AE-based models: AE + Supervised Multi-Layer Perceptron (sMLP) and End-to-End Supervised AE (EEsAE).
• Utilized diagnostic codes from one million patients in the Korean National Health Information Database.
• Compared performance against baseline models: eXtreme Gradient Boosting (XGB) and Naïve Bayes.

Main Results:

• The EEsAE model achieved the highest F1-score (0.86) and area under the curve.
• EEsAE and AE + sMLP demonstrated the highest recall rates.
• XGB model showed the highest precision; key influential diagnoses identified.

Conclusions:

• The novel EEsAE model demonstrates significant potential for predicting disease co-existence.
• This approach highlights the utility of deep learning with diagnostic codes for clinical applications.