A maintenance hemodialysis mortality prediction model based on anomaly detection using longitudinal hemodialysis data

Area of Science:

• Medical Informatics
• Machine Learning in Healthcare
• Renal Disease Management

Background:

• End-stage renal disease patients on hemodialysis (HD) face high mortality and financial burdens.
• Healthcare systems, particularly in China, experience significant patient follow-up losses, complicating mortality risk identification.
• Existing methods struggle with imbalanced datasets common in longitudinal health records.

Purpose of the Study:

• To propose a hemodialysis mortality prediction approach using longitudinal HD data.
• To address data imbalance issues caused by patient follow-up losses.
• To identify patients with high mortality risks in maintenance HD.

Main Methods:

• A long short-term memory autoencoder (LSTM AE) model was developed to track patient condition changes.
• Anomaly detection theory was applied, training on surviving patients and identifying non-survivors via reconstruction errors.
• Model performance was benchmarked against logistic regression, SVM, random forest, LSTM classifier, isolation forest, and stacked autoencoder models.

Main Results:

• The LSTM AE model demonstrated superior performance in predicting mortality, outperforming other models across various prediction windows.
• Key performance metrics included an Area Under the PR Curve of 0.57, Recall_macro of 0.86, and F1-score_macro of 0.87.
• Dialysis session length was identified as the most significant predictor in the mortality risk model.

Conclusions:

• The proposed LSTM AE approach effectively detects high-risk hemodialysis patients from imbalanced datasets.
• Anomaly detection theory provides a viable strategy for mortality prediction with longitudinal health data.
• This method offers a promising tool for improving patient management and reducing mortality in maintenance HD populations.