Clinical Context Is More Important than Data Quantity to the Performance of an Artificial Intelligence-Based Early...

Area of Science:

• Clinical Informatics
• Artificial Intelligence in Healthcare
• Predictive Analytics

Background:

• Clinical data quantity varies, influenced by clinician risk perception and testing decisions.
• Missing electronic health record data can signify that tests were deemed unnecessary for stable patients.

Purpose of the Study:

• To evaluate a deep learning early warning system (VitalCare-Major Adverse Event Score) for predicting critical events in adult inpatients.
• To assess if patient severity (Charlson Comorbidity Index - CCI) and laboratory data volume impact predictive accuracy.

Main Methods:

• Retrospective single-center study of adult inpatients.
• Utilized a deep learning model to predict unplanned ICU transfers, cardiac arrests, or death 6 hours in advance.
• Classified patients by CCI and analyzed the impact of data completeness and volume.

Main Results:

• Patients with high CCI had more testing and less missing data; moderate-to-low CCI patients had less testing and more missing data.
• Despite differences in data quantity, missingness patterns were similar in event cohorts across CCI groups.
• The deep learning model's predictive performance was robust across severity groups, indicating clinical context's importance.

Conclusions:

• Data completeness in electronic health records requires a nuanced interpretation.
• Preserving real-world patterns of laboratory test ordering can improve predictive model performance.
• The clinical context of missing data is a critical factor in predictive accuracy.