Early Detection of Post-Surgical Complications using Time-series Electronic Health Records

Area of Science:

• Medical Informatics
• Machine Learning in Healthcare
• Time-Series Analysis

Background:

• Electronic health records (EHR) present challenges for predictive modeling due to asynchronous and missing data.
• Traditional methods often require data truncation or heuristic imputation, potentially introducing bias and reducing model performance.
• Accurately capturing the longitudinal nature of patient health status is crucial for reliable complication prediction.

Purpose of the Study:

• To develop an augmented gated recurrent unit (GRU) model that integrates missingness and timeline signals directly into its architecture.
• To address the limitations of conventional data preprocessing techniques in handling complex EHR data.
• To improve the prediction of post-operative bleeding (POB) using real-world patient data.

Main Methods:

• An augmented gated recurrent unit (GRU) model was designed to incorporate both data missingness and temporal information.
• The model was evaluated using a dataset of post-operative bleeding (POB) from the Mayo Clinic EHR system.
• Conventional models were trained on heuristically imputed datasets for comparative analysis.

Main Results:

• The proposed augmented GRU model demonstrated superior performance compared to state-of-the-art methods in detecting post-operative bleeding (POB).
• The model's ability to handle asynchronous and missing data in EHR was validated.
• Performance improvements indicate greater eligibility for real-world EHR data analysis.

Conclusions:

• The augmented GRU model offers a robust solution for analyzing longitudinal EHR data, overcoming limitations of traditional approaches.
• This approach enhances the accuracy of predicting critical health events like post-operative bleeding (POB).
• The model shows significant potential for clinical decision support systems utilizing EHR data.