Deep STI: Deep Stochastic Time-series Imputation on Electronic Health Records.

Area of Science:

• Healthcare Analytics
• Deep Learning
• Medical Informatics

Background:

• Electronic Health Records (EHRs) are vital for healthcare analytics but contain missing values that impede deep learning model performance.
• Existing missing data imputation methods struggle to capture the temporal dynamics crucial for EHR data analysis.
• Accurate imputation is essential for reliable clinical decision-making and patient care.

Purpose of the Study:

• To introduce the Deep Stochastic Time-series Imputation (Deep STI) algorithm, an end-to-end deep learning model for handling missing EHR data.
• To leverage temporal dynamics in EHRs for accurate inference of missing values.
• To improve predictive accuracy for diseases like liver cancer using imputed EHR data.

Main Methods:

• Developed Deep STI, integrating a sequence-to-sequence generative network with a prediction network.
• Trained and evaluated Deep STI on liver cancer data from the National Taiwan University Hospital (NTUH).
• Compared Deep STI's performance against extreme gradient boosting and Transformer models.

Main Results:

• Deep STI achieved superior 5-year hepatocellular carcinoma prediction accuracy (19.21% AUC) compared to extreme gradient boosting (18.15%) and Transformer (18.09%).
• Ablation studies confirmed the effectiveness of the generative architecture in imputation.
• The model demonstrated high accuracy in inferring missing values by leveraging temporal context.

Conclusions:

• Deep STI significantly enhances the reliability of disease analysis using incomplete EHR data.
• The algorithm sets a new standard for predictive healthcare analytics utilizing deep learning on EHRs.
• This work advances healthcare analytics and promotes further research in deep learning applications for EHR data.