A classification framework for exploiting sparse multi-variate temporal features with application to adverse drug...

Area of Science:

• Health Informatics
• Machine Learning
• Clinical Decision Support

Background:

• Adverse drug events (ADEs) and preventable medical errors cost the US healthcare system billions annually.
• Reducing ADEs is crucial for patient well-being and economic sustainability in healthcare.
• Existing predictive models often overlook temporal data in electronic health records (EHRs).

Purpose of the Study:

• To develop a novel classification framework for detecting ADEs in EHRs.
• To leverage the temporality and sparsity of features within EHR data for improved prediction.
• To create a versatile and computationally efficient method for ADE detection.

Main Methods:

• A three-phase framework transforms sparse, multi-variate time-series EHR features into a usable representation.
• Three strategies are proposed and evaluated for incorporating feature sparsity into the new representation.
• The framework is designed to be compatible with various classification algorithms.

Main Results:

• The proposed framework significantly outperforms state-of-the-art methods on 15 real-world ADE datasets.
• The model successfully identified clinically relevant features associated with ADE detection.
• Evaluations were conducted on a large-scale, real-world EHR system.

Conclusions:

• Temporal, multi-variate, and sparse features from EHRs are effective for predicting ADEs.
• The framework is method-agnostic and computationally inexpensive, making it broadly applicable.
• This work provides a valuable foundation for future machine learning applications in EHR analysis.