Summary
- A study predicting weight changes in heart failure patients using EMRs was conducted.
- The study used the EMRs database of AMC, Seoul, South Korea, between January 2000 and November 2021, ensuring data accuracy and completeness through a de-identification system.
- A novel LSTM framework was designed to predict weight changes in heart failure patients treated with loop diuretics using EMRs, involving data preparation, model experiments, and comparison of five AI models.
- Cohort selection, feature preparation, outlier detection, and data resampling were carried out to optimize the model, involving 65 variables from medications, lab tests, vital signs, demographics, and diagnosis.
- The final LSTM with attention mechanisms model named TSFD-LSTM was selected as the final model based on the lowest error rate, demonstrating the potential for predicting weight changes in heart failure patients within a 1 kg margin of error.
Researchers at AMC obtained ethical approval for a study investigating weight changes in heart failure patients treated with loop diuretics using electronic medical records (EMRs). The study utilized a novel LSTM framework and excluded the need for informed consent as the data analyzed was anonymized. The data source was the EMRs database of AMC in South Korea, and the study process involved data preparation, model experiments, and final model selection.
Patients meeting specific inclusion criteria were selected for the study, and various medical variables were extracted from the EMRs for feature preparation. Outlier detection and data normalization were performed, and data resampling into time-series sequences was carried out. The data was split into training, testing, and validation datasets, and five AI models were developed and compared for performance using metrics such as mean absolute error and root mean squared error.
The study utilized a sequence-to-sequence learning approach with a focus on Long short-term memory (LSTM) and attention mechanisms to predict weight changes accurately in heart failure patients. The LSTM model addressed the vanishing gradient problem of conventional recurrent neural networks, and the attention mechanism enhanced model performance by attending to important parts of input sequences during prediction. The final model, TSFD-LSTM, was selected based on its low prediction errors and predictive accuracy within 1 kg, making it suitable for effective diuretic dose adjustments and treatment planning in clinical settings.
Internal Medicine, Nephrology, Pharmacists