1188 Test Characteristics of a Machine Learned Electronic Medical Record Extractable Tool for OSA Case Identification in a Community-Based Population. (27th May 2020)

• Record Type: Journal Article
• Authors: Patel, S I; Kukafka, D; Antonescu, C; Combs, D; Lee-Iannotti, J; Quan, S F; Parthasarathy, S
• Is Part Of: Sleep. Volume 43(2020)Supplement 1
• Page Start: A454
• Page End: A454
• Publication Date: 2020-05-27
• Subjects: Sleep -- Physiological aspects -- Periodicals; Sleep disorders -- Periodicals; Sommeil -- Aspect physiologique -- Périodiques; Sommeil, Troubles du -- Périodiques; Sleep disorders; Sleep -- Physiological aspects; Sleep -- physiological aspects; Sleep Wake Disorders; Psychophysiology; Electronic journals; Periodicals; 616.8498
• DOI: 10.1093/sleep/zsaa056.1182
• Languages: English
• ISSNs: 0161-8105

Abstract

Abstract: Introduction: Obstructive sleep apnea (OSA) is a significantly underdiagnosed medical condition. A machine learning method known as SLIM (Supersparse Linear Integer Models) that can be extracted from the Electronic Health Record (EHR) has found to be superior to patient-reported sleep-related symptoms to diagnose OSA. Such an evaluation, however, was previously validated in a laboratory-based population. Our aim was to determine the test characteristics for the EHR-extractable SLIM tool in a community-based population. Methods: Subjects who participated in the Sleep Heart Health Study (SHHS) were included in this analysis. Variable definitions of OSA were determined using an Apnea Hypopnea Index (AHI) threshold of 5 per hour, 15 per hour, or the presence of any comorbidity (hypertension, ischemic heart disease, stroke, mood disorders, impairment of cognition, or sleepiness) when the AHI was between 5 to 15 per hour. Variable hypopnea definitions based upon degree of oxygen desaturation and associated arousals were considered. Results: In the SHHS dataset, the Receiver Operating Characteristics (ROC) for a SLIM score threshold of 9 for men and 5 for women was good when OSA was defined by AHI > 5 per hour (hypopneas with either > 3% oxygen desaturation or arousals). Specifically, the ROC was 0.72 (95% Confidence Intervals [CI] 0.70; 0.74) with a Positive Predictive Value [PPV] of 0.98 and Likelihood Ratio of a positive test (LR+) of 11.3. The LR+ (6.0) and PPV (0.92) were also good when an AHI of 5 per hour threshold was adopted with hypopneas scored using the minimum 3% oxygen desaturation alone. Similarly, the ROC was good 0.74 (95%CI 0.73; 0.76) with a Positive Predictive Value [PPV] of 0.98 and Likelihood Ratio of a positive test (LR+) of 11.3. The LR+ (8.9) and PPV (0.81) were also good in the presence of comorbidities when AHI was 5 to 15 per hour using > 4% oxygen desaturation alone. Conclusion: The EHR-extractable tool can be an actionable tool for case-identification of patients needing a referral for sleep study in a community-based population. Such an approach could facilitate an automated, rather than manual, OSA screening approach aimed at managing population health. Support: HL138377