Detecting anomalous sequences in electronic health records using higher-order tensor networks.

Area of Science:

• Health Informatics
• Data Science
• Machine Learning

Background:

• Electronic health records (EHR) generate vast patient data vital for improving healthcare.
• Detecting anomalous sequences in EHR is challenging due to data imbalance, complex event relationships, and high dimensionality.
• Conventional methods often overlook event details and higher-order dependencies, limiting sequence discrimination.

Purpose of the Study:

• To develop novel algorithms for detecting anomalous event sequences and subsequences in EHR.
• To address limitations of existing methods by incorporating network-based representations, variable higher-order dependencies, and event durations.
• To improve the accuracy and robustness of anomaly detection in healthcare data.

Main Methods:

• Proposed algorithms utilize network-based representations of event interactions.
• Algorithms account for variable higher-order dependencies and incorporate event durations.
• Anomaly detection is performed by quantifying graph changes after sequence removal using graph distance metrics.

Main Results:

• The proposed event sequence anomaly detection algorithm demonstrated superior performance compared to baseline methods.
• Effectiveness was validated on both synthetic datasets and real-world EHR data.
• The subsequence algorithm provides plausible path recommendations and salient information for detected anomalies.

Conclusions:

• The novel network-based approach effectively detects anomalous sequences in EHR data.
• Incorporating event durations and higher-order dependencies enhances anomaly discrimination.
• This research offers a promising advancement for securing health information technology systems and improving patient care.