Construction of Risk Prediction Model of Type 2 Diabetic Kidney Disease Based on Deep Learning

Area of Science:

• Biomedical Informatics
• Machine Learning in Healthcare
• Nephrology Research

Background:

• Diabetic kidney disease (DKD) poses a significant health burden.
• Accurate prediction of DKD is crucial for timely intervention.
• Existing prediction models may not fully capture disease dynamics.

Purpose of the Study:

• To develop and evaluate a DKD prediction model utilizing Long Short-Term Memory (LSTM) neural networks.
• To assess the impact of key physiological variabilities on DKD prediction accuracy.
• To establish a robust model for early identification of patients at risk of DKD.

Main Methods:

• Literature review and physician focus groups identified DKD risk factors.
• A Long Short-Term Memory (LSTM) neural network was developed using Pytorch.
• Data from 6,040 type 2 diabetes mellitus patients over 7 years were analyzed.
• Model performance was evaluated using accuracy, precision, recall, and Area Under the Curve (AUC) of the Receiver Operating Characteristic (ROC) curve.
• The influence of glycosylated hemoglobin (HbA1c), systolic blood pressure (SBP), and pulse pressure (PP) variabilities was investigated.

Main Results:

• The LSTM-based DKD prediction model achieved an accuracy of 83% and an AUC of 0.83.
• Removing HbA1c, SBP, or PP variability significantly decreased model accuracy (to 78%, 79%, and 81% respectively, P<0.001).
• Excluding these variabilities also substantially reduced AUC values (to 0.72, 0.75, and 0.77 respectively, P<0.05).

Conclusions:

• The developed LSTM neural network model demonstrates high accuracy and AUC for DKD risk prediction.
• Incorporating HbA1c, SBP, and PP variabilities as features significantly enhances model performance.
• This approach offers a promising tool for improving early detection and management of DKD.