Predictability Bounds of Electronic Health Records

Area of Science:

• Health Informatics
• Computational Medicine
• Information Theory

Background:

• Advancing healthcare delivery and reducing costs are critical societal challenges.
• Predicting disease progression from patient medical history is essential for effective intervention.
• Electronic Health Records (EHRs) offer vast data for studying disease patterns.

Purpose of the Study:

• To quantify the inherent predictability of disease progression using an information-theoretic methodology.
• To analyze disease history predictability in a large-scale electronic health records dataset.
• To investigate the impact of individual versus collective disease histories and temporal dependencies on predictability.

Main Methods:

• Utilized an information-theoretic framework to assess disease history predictability.
• Analyzed a large electronic health records dataset comprising over half a million patients.
• Examined predictability from zeroth-order statistics to temporally informed statistics, considering individual and collective patient data.

Main Results:

• Common disease progression patterns significantly increase predictability compared to independent disease histories.
• Identified the order at which temporal dependence structures diminish with increasing time-correlated statistics.
• Shuffling individual disease histories resulted in only a marginal decrease in predictability bounds.

Conclusions:

• While common disease sequences enhance predictive accuracy, the importance of strict temporal ordering in patient histories may be less than anticipated.
• The findings highlight the complexity of healthcare processes and challenges in developing universal prediction algorithms.
• Further research is needed to fully leverage EHR data for accurate and reliable disease prediction.