By Justin Schaper and Megan Bultema
Artificial intelligence (AI) is revolutionizing healthcare data analytics and changing the way we predict, learn and act based on insights gained through AI-powered data models. At Olive, we’ve learned that solving healthcare problems requires multiple types of solutions and technologies. In addition to AI, we use natural language processing (NLP), machine learning (ML) and expert systems to identify and address unwarranted clinical variation in surgery.
The example of cohorts
Olive uses a combination of NLP, expert systems and ML models to create and maintain “cohorts.” Cohorts are a method for grouping surgical cases into clinically similar groupings that are more meaningful to surgeons and staff than the traditional groupings based on billing and reimbursement codes, such as DRGs, CPTs and ICD-10s.
To establish these groupings, Olive clinicians developed a library of cohort definitions using a rules-based approach that looked at both discrete data points and concepts abstracted via NLP from the clinical operative report. These rules and definitions form a knowledge base for an Olive-designed expert system. The combination of NLP and expert system technologies creates a refined, clinically meaningful set of high-integrity data that enable the quick identification and root cause analysis of unwarranted clinical variance.
Healthcare data is complex and inconsistent
Healthcare data is not consistent across healthcare systems, hospitals, surgeons and EHRs. Surgeons may use slightly different language in their operative reports than their colleagues. EHRs may capture data points in slightly different ways. Coders may encode billing information differently. These sorts of discrepancies can challenge traditional expert systems that rely on rules created and defined to expect a static set of possible input values. To maintain the integrity of the cohorts, Olive has applied ML models to work in conjunction with the cohort engine.
Combining machine learning and expert systems
The ML models were trained from a broad set of data to recognize patterns and make predictions about likely cohort placement for a given surgical encounter. These predictions are done independently from the cohort engine. By comparing the ML predictions to the cohort engine output, Olive can rapidly find where potential cohort rule specific data anomalies cause invalid cohort placement. By combining these approaches, Olive is able to quickly validate and provide ongoing monitoring of client data to ensure that surgical cases are in the appropriate grouping and that the integrity of the unwarranted variance analysis is supported.