Development and evaluation of a machine learning algorithm to map medical conditions and procedures from real-world data

Background: Ontologies characterize complex and detailed information and are extensively used in healthcare research. Medical information (textbooks, expert opinions, clinical evidence) has information on conditions and its corresponding procedures (treatments), but this information is not captured or structured in any ontology. The objective of the research is to create a condition-procedure ontology from real world data to be utilized in observational research or electronic health record (EHR) system. Methods: Predictive models are developed to learn from five datasets (administrative claims, hospital charge data) to generate two algorithms (diagnostic and therapeutic) to predict condition-procedure relationships in the SNOMED-CT vocabulary. A reference set with 100 positive pairs per algorithm, and 32,132 negative pairs were developed. Predictive models were constructed by designing 51 possible covariates that describe condition-procedure pairs from Optum© De-Identified Clinformatics® Data Mart Database – Socio-Economic Status (Optum) dataset and determining which covariates discriminated between the positive and negative controls, as measured by Area Under Receiver Operator Characteristic Curve (AUC). External validation of the final algorithms was performed on 4 other databases. The final algorithms were applied across the universe of condition and procedure pairs in all five databases to construct the full condition-procedure ontology, and the ontology was evaluated for validity and coverage of condition and procedure concepts from the set of identified condition-procedure pairs. An additional analysis was trained to classify diagnostic vs. therapeutic intervention based on the overlap of pairs within the two algorithms. Results: Algorithms include the following covariates: condition-procedure occurring together, relative risk, support and sensitivity. Both algorithms had AUCs greater than .90, and external validation also showed similar results. In Optum, 98% of conditions and 63% of procedure codes had at least one relationship identified in the ontology. The intervention type analysis resulted in an AUC of 0.79. Conclusions: Real-world data can be utilized to construct a medical ontology of condition-procedure relationships with strong performance and good coverage. These results can be utilized to fuel research efforts in healthcare such as cohort generation and computer provider order entry systems by understanding conditions and procedures and their application to diagnose or treat a patient.Ph.D.Includes bibliographical reference