0317 COMPARISON OF TWO VALIDATION TECHNIQUES USING THE ALLIANCE SLEEP QUESTIONNAIRE (ASQ) INSOMNIA MODULE

The Alliance Sleep Questionnaire (ASQ) is a comprehensive, on-line sleep questionnaire containing ~600 variables in multiple formats, including questions where more than one answer can be selected. We analyzed ASQ data using regression and machine learning techniques to identify Stanford Sleep Disorders Clinic (SSDC) insomnia patients. The population included SSDC patients who completed the ASQ and signed consent. Patients were considered positive for insomnia if they scheduled a Behavioral Sleep Medicine Program appointment. Remaining patients were considered negative for insomnia, as most SSDC patients are seen for OSA. For the machine learning approach, we analyzed 154 continuous and categorical variables from the ASQ related to insomnia. Sequential forward feature selection was applied to identify a subset of variables. Bagged decision trees were used as a classification model to accommodate missing values and categorical data. Two regression models were created. The first approach included 15 potential variables selected using clinical judgement. The second, hybrid model, included the initial 15 variables plus 4 additional variables identified in the machine learning approach and considered clinically relevant. Backward elimination stepwise selection was used to build the regression models. Performance was evaluated using sensitivity, specificity, positive predictive value (PPV), negative predictive value (NPV), and area under the receiver operator curve (AUC). Optimal diagnosis point threshold was selected by weighting sensitivity and specificity equally. 5701 patients (3239 males, 2462 females) were analyzed and 944 met insomnia criteria. Machine learning selected 35 features using a learning dataset, resulting in Sensitivity=73.2%, Specificity=73.1%, PPV=35.0%, NPV=93.3%, and AUC=0.80 when applied to the test data. The original ASQ model included 8 covariates with Sensitivity=71.1%, Specificity=69.6%, PPV=29.5%, NPV=92.8%, and AUC=0.78. The hybrid model contained 13 covariates with Sensitivity=75.1%, Specificity=73.0%, PPV=35.7%, NPV=93.6%, and AUC=0.81. Although the machine learning approach had slightly higher specificity and sensitivity compared to the original regression model, it took longer to build and process the data. The hybrid regression model, with covariates selected using a combination of clinical judgement and machine learning, had the best overall performance. Philips Respironics Foundation grant, the Stanford Center for Sleep Sciences and Medicine, and gift funds.

Fetched live from OpenAlex and de-inverted. Abstracts are not stored in this database: the inverted indexes are 8.6 GB of the frame’s 9.3 GB of text, and the host has 13 GB free.