PheValuator

PheValuator estimates sensitivity, specificity, positive predictive value (PPV), and negative predictive value (NPV) of phenotype algorithms in observational health databases by building probabilistic diagnostic predictive models.

Key Features:

• Diagnostic Predictive Modeling: Employs machine learning to create diagnostic predictive models that generate probabilistic gold-standard phenotype probabilities.
• Comprehensive Evaluation Across Datasets: Has been applied to administrative claims datasets including OptumInsight's de-identified Clinformatics™ Datamart and IBM MarketScan databases.
• Algorithm Comparison and Performance Metrics: Compares predicted phenotype probabilities against phenotype algorithm inclusion/exclusion criteria and evaluates configurations such as 1X and 1X-IP-1stPos to derive sensitivity, specificity, PPV, and NPV.
• Enhanced Predictive Capabilities (PheValuator 2.0): Incorporates diagnostic conditions, clinical observations, drug prescriptions, laboratory measurements, and temporal relationships of predictors to improve predictive accuracy.
• Improved Accuracy and Reliability: Produces performance estimates that align more closely with traditional validation methods such as expert chart review, reducing discrepancies in PPV, sensitivity, and specificity.

Scientific Applications:

• Phenotype Algorithm Evaluation: Quantitatively assesses the performance of rule-based phenotype algorithms in observational health data.
• Reproducibility and Reliability Assessment: Supports evaluation of reproducibility and reliability of phenotype definitions across healthcare datasets.
• Quantitative Bias Analysis Support: Provides performance estimates that can be used in quantitative bias analysis for observational studies.
• Cross-dataset Validation: Facilitates comparison and validation of disease phenotype performance across administrative claims datasets.

Methodology:

Uses machine learning to build diagnostic predictive models that produce probabilistic phenotype probabilities from diagnostic conditions, clinical observations, drug prescriptions, laboratory measurements, and their temporal relationships, then compares these probabilities to phenotype algorithm criteria to compute sensitivity, specificity, PPV, and NPV; evaluations have been performed on OptumInsight Clinformatics™ and IBM MarketScan datasets.