Probabilistic techniques for obtaining accurate patient counts in Clinical Data Warehouses

Area of Science:

• Computational biology
• Health informatics
• Biostatistics

Background:

• Accurate patient counts are essential for clinical trials, quality measures, and medical research.
• Existing Clinical Data Warehouses (CDWs) often suffer from incomplete or inaccurate patient data.
• Traditional methods for patient counting may not yield reliable results.

Purpose of the Study:

• To explore the application of probabilistic techniques for obtaining accurate patient counts from a synthetic Clinical Data Warehouse.
• To evaluate and compare different probabilistic methods against traditional counting techniques.
• To demonstrate the utility of Bayesian approaches in improving patient count accuracy.

Main Methods:

• Developed a synthetic Clinical Data Warehouse with a custom patient data generation engine.
• Modeled medical billing as a diagnostic test for condition presence, calculating sensitivity and specificity.
• Implemented a "Bayesian Chain" approach using Bayes' Theorem for sequential probability updates and a "one-shot" approach based on any billing occurrence.
• Conducted a "Simulated Expert Review" for ground truth validation.

Main Results:

• Probabilistic approaches significantly improved patient count accuracy compared to raw counts.
• The "Simulated Expert Review" combined with a single application of Bayes' Theorem yielded the best performance with an average error rate of 2.1%.
• Straightforward billing counts resulted in a substantially higher average error rate of 43.7%.

Conclusions:

• Bayesian probabilistic approaches offer a substantial improvement in patient count accuracy for simulated patient populations.
• These techniques are valuable for applications requiring precise patient metrics, such as clinical trial management and quality assessment.
• The proposed Bayesian framework has broad applicability for enhancing data accuracy in healthcare.