Learning temporal weights of clinical events using variable importance.

Area of Science:

• Health Informatics
• Machine Learning
• Pharmacovigilance

Background:

• Electronic health records (EHRs) are valuable for monitoring adverse drug events (ADEs), but ADEs are under-reported.
• Machine learning can automatically detect potential ADEs from EHR data.
• Incorporating the temporality of clinical events is crucial for accurate ADE detection.

Purpose of the Study:

• To develop and evaluate methods for learning temporal weights of clinical events for improved ADE detection.
• To compare the effectiveness of learned temporal weights against pre-assigned weights.
• To investigate the impact of different weight application strategies and granularity levels.

Main Methods:

• Utilized random forest variable importance to derive temporal weights for clinical events.
• Proposed two strategies for applying learned weights: weighted aggregation and weighted sampling.
• Compared the predictive performance of models using learned weights versus pre-assigned weights.

Main Results:

• The weighted sampling strategy significantly improved predictive performance when using learned temporal weights compared to pre-assigned weights.
• No significant performance difference was observed between learned and pre-assigned weights in the weighted aggregation strategy.
• The granularity of weight learning impacted the weighted sampling strategy but not weighted aggregation.

Conclusions:

• Learning temporal weights is significantly beneficial for predictive performance in ADE detection, particularly with the weighted sampling strategy.
• Weighted aggregation generally reduces the impact of temporal weighting, regardless of whether weights are learned or pre-assigned.