Potential of risk-based population guidelines to reduce cardiovascular risk in a large integrated health system.

OBJECTIVES: We evaluated an alternative way to implement guidelines using an automated risk calculator and risk-based decision tool to calculate patients' risk of cardiovascular disease (CVD) and recommend therapies. We compared such an approach with traditional guidelines. STUDY DESIGN: A retrospective cohort study of 1,506,109 Kaiser Permanente Southern California members 35 years or older. METHODS: We estimated 3-year risks of fatal and nonfatal myocardial infarction and stroke using an independently developed risk calculator, then graphically compared risks with observed outcomes. We used the area under the receiver operating characteristics curve to assess discrimination, and the Hosmer-Lemeshow statistic to test fit. We compared the characteristics and outcomes of populations identified for medication therapy by the risk-based decision tool and traditional guidelines using bivariate statistics. RESULTS: A risk score was obtained in 72% (1,082,158) of members. The risk calculator was fairly good in discrimination: the area under the curve was 0.774 (95% CI, 0.770-0.779) for myocardial infarction and 0.805 (95% CI, 0.801-0.808) for stroke. Predictiveness and fit was good based on graphical analysis and Hosmer-Lemeshow P < .0001. The risk-based decision tool identified high-risk patients for treatment who were not identified by traditional guidelines (3.80% of all those identified for statins, 3.04% for antihypertensives), as well as low-risk patients who were identified by guidelines (3.80% for statins, 2.51% for antihypertensives). CONCLUSIONS: The risk calculator provided risk estimates in most patients and demonstrated fairly good discrimination and predictiveness. The risk-based decision tool identified high-risk patients for treatment not identified by traditional guidelines, as well as low-risk patients for whom treatment may be unnecessary.

Automated identification of patients with pulmonary nodules in an integrated health system using administrative health plan data, radiology reports, and natural language processing.

INTRODUCTION: Lung nodules are commonly encountered in clinical practice, yet little is known about their management in community settings. An automated method for identifying patients with lung nodules would greatly facilitate research in this area. METHODS: Using members of a large, community-based health plan from 2006 to 2010, we developed a method to identify patients with lung nodules, by combining five diagnostic codes, four procedural codes, and a natural language processing algorithm that performed free text searches of radiology transcripts. An experienced pulmonologist reviewed a random sample of 116 radiology transcripts, providing a reference standard for the natural language processing algorithm. RESULTS: With the use of an automated method, we identified 7112 unique members as having one or more incident lung nodules. The mean age of the patients was 65 years (standard deviation 14 years). There were slightly more women (54%) than men, and Hispanics and non-whites comprised 45% of the lung nodule cohort. Thirty-six percent were never smokers whereas 11% were current smokers. Fourteen percent of the patients were subsequently diagnosed with lung cancer. The sensitivity and specificity of the natural language processing algorithm for identifying the presence of lung nodules were 96% and 86%, respectively, compared with clinician review. Among the true positive transcripts in the validation sample, only 35% were solitary and unaccompanied by one or more associated findings, and 56% measured 8 to 30 mm in diameter. CONCLUSIONS: A combination of diagnostic codes, procedural codes, and a natural language processing algorithm for free text searching of radiology reports can accurately and efficiently identify patients with incident lung nodules, many of whom are subsequently diagnosed with lung cancer.