2019-997 Use of Machine Learning to Predict Non-Diagnostic Home Sleep Apnea Tests
Web Published:
11/30/2019
SUMMARY:
Researchers led by Robert Stretch from the Division of Pulmonary, Critical Care & Sleep Medicine at UCLA have developed an algorithm that can predict whether a patient will have a non-diagnostic home sleep apnea test based upon data from the electronic health record and a brief questionnaire.
BACKGROUND:
Obstructive sleep apnea (OSA) affects between 4-37% of the adult population depending on the diagnostic criteria applied and population studied. Diagnostic testing for OSA typically starts with an “unattended” home sleep apnea test (HSAT) using a portable device. Since this test has a 17% false-negative rate, it is recommended that all patients who have a non-diagnostic initial HSAT should undergo an “attended” in-laboratory polysomnogram (PSG). A non-diagnostic HSAT is one in which the recording is technically inadequate (i.e. due to signal loss) or appears normal (i.e. respiratory event index < 5/hr). In clinical practice the rate of non-diagnostic HSATs varies between 15-30% of all studies. The ability to predict a non-diagnostic HSAT result prior to the test being ordered allows clinicians to pre-emptively order a PSG instead, thereby minimizing harms in the form of delayed diagnoses, missed diagnoses, additional financial burden to the patient and healthcare system, and inefficient use of limited resources.
INNOVATION:
Researchers led by Robert Stretch from the Division of Pulmonary, Critical Care & Sleep Medicine at UCLA have developed an algorithm that predicts whether a patient will have a non-diagnostic home sleep apnea test. The algorithm was developed using machine learning techniques and uses data from the electronic health record (automatically sourced) and patient responses to a brief questionnaire to make predictions with a high degree of precision and accuracy. Assuming all patients with a non-diagnostic HSAT subsequently undergo PSG (as per American Academy of Sleep Medicine guidelines), implementing this model to guide testing would result in the following:
For every 1000 patients currently undergoing HSAT as their initial test for OSA…
- The number of patients needing to undergo both HSAT and PSG before obtaining a diagnosis would decrease by 139 (45.9%). These patients would instead be diagnosed on the first test.
- The absolute diagnostic yield of initial sleep studies would increase by 13.9% (83.6% up from 69.7%).
- The absolute diagnostic yield of HSATs would increase by 10.5% (80.2% up from 69.7%).
- 172 fewer HSATs (17.2% decrease) and only 33 additional PSGs (10.9% increase) would be performed. This represents a net cost saving based on published literature regarding the relative cost of HSAT and PSG.
- The algorithm’s classification threshold can also be adjusted to meet the specific needs of an institution. For example, one such alternative threshold results in a 24.4% reduction in the number of patients needing both HSAT and PSG while only increasing total PSGs performed by 2.6%.
POTENTIAL APPLICATIONS:
- Predicting non-diagnostic HSAT results
- Integration into sleep study triage processes in hospitals and clinics
- Use by insurers to optimize sleep study prior authorization
ADVANTAGES:
- Unique approach that improves the economic and diagnostic efficiency of HSATs while minimizing the harms associated with non-diagnostic studies.
- Optimizes the allocation of limited sleep resources (HSAT devices, PSG beds) required for the diagnosis of OSA.
- Patient-centered approach to testing that reduces average time-to-diagnosis and improves patient satisfaction.
RELATED MATERIALS:
- Robert Stretch, Abigail Maller, Dennis Hwang, Armand Ryden, Constance Fung, Michelle Zeidler, 0465 Performance of Revised Machine Learning Models for Prediction of Non-Diagnostic Home Sleep Apnea Tests, Sleep, Volume 42, Issue Supplement_1, April 2019, Page A187, https://doi.org/10.1093/sleep/zsz067.464
DEVELOPMENT-TO-DATE:
Initial derivation and validation in a split cohort of 613 patients.
Patent Information:
Title |
App Type |
Country |
Serial No. |
Patent No. |
File Date |
Issued Date |
Expire Date |