Utilizing time series data embedded in electronic health records to develop continuous mortality risk prediction...

Area of Science:

• Medical Informatics
• Clinical Decision Support
• Machine Learning in Healthcare

Background:

• Continuous mortality risk monitoring is crucial for patient management and resource allocation in hospitals.
• Electronic health records (EHR) present challenges for continuous risk prediction due to data incompleteness and irregularity.

Purpose of the Study:

• To propose and validate a novel framework for continuous mortality risk monitoring using EHR data.
• To compare the performance of a temporal prediction model against several non-temporal machine learning methods.

Main Methods:

• Utilized hidden Markov models (HMMs) as a temporal technique, incorporating patient's prior health states and current clinical data.
• Applied the framework to 3898 patient encounters with suspected infection, adhering to Sepsis-3 criteria.
• Evaluated performance using Area Under the Receiver Operating Characteristics (AUROC) curve, sensitivity, specificity, and G-mean.

Main Results:

• The proposed temporal HMM framework achieved an AUROC of 0.87, outperforming non-temporal methods (DT, NB, SVM, LR, RF) by 9-12%.
• The model demonstrated a superior G-mean of 0.78, indicating a better balance between sensitivity and specificity than existing bedside criteria (G-mean=0.71).
• The framework effectively leverages longitudinal EHR data for improved predictive performance.

Conclusions:

• The developed temporal framework offers a significant advancement in continuous mortality risk prediction from EHR data.
• This approach enhances the ability to identify changes in patient health status over time, facilitating timely clinical interventions and treatment strategies.
• The findings suggest potential for improved patient outcomes and optimized hospital resource utilization through proactive risk management.