Using AI to Transform CPAP Compliance

AI's Impact on Personalization and Value Creation

Artificial Intelligence (AI) is one of the most prevalent and transformative technologies available today. In addition to its visibility in the tech sector, AI has been strategically integrated across businesses as diverse as Netflix, Spotify, The New Yorker Magazine and the BBC World News. What these and many other companies have learned is that AI’s effective personalization increases the usage of a product or service by customizing the user experience. Ultimately using AI leads to greater customer satisfaction and creates value. AI can be most transformative in its ability to gain insight through data analysis and its use in healthcare and sleep medicine in particular is breaking new ground. A recent BetterNight study, done in collaboration with EnsoData, had one of the largest CPAP data sets to date and proved that AI can be used for more than diagnosis or treatment and by analyzing behavior and usage types, can be predictive of CPAP compliance.

Addressing CPAP Non-Adherence: A Catalyst for AI in Sleep Medicine

The catalyst driving the use of AI in sleep medicine is the extremely high non-adherence rate in CPAP. Recently published data from a study in the Journal of Otolaryngology – Head and Neck Surgery revealed a staggering 34.1% CPAP non-adherence rate. Given that 18% of the US population suffers from OSA, this non-adherence rate adds up to millions of patients who are needlessly suffering. The first step in BetterNight’s research was to look at CPAP adoption and management through the lens of behavior change using the transtheoretical model. This model divides the behavior change journey into 5 key stages. The first stage is pre-contemplation, where change has not yet been considered. The next stage is contemplation, where there is awareness of an issue and momentum starts to build. In the determination stage, change is put into motion. In the action stage, the new behavior is occurring while in the maintenance stage the new behavior is established and the patient wants it to continue. Even though a patient may be determined to succeed, they will slip on occasion opening up the possibility of falling back into earlier behavioral stages. The further back they fall, the more difficult it is to get them to the maintenance stage. The goal in this study as it relates to CPAP adherence is early intervention, ideally getting to patients before they slip back initially, and before they relapse any further.

Analyzing CPAP Usage Patterns: Insights from Behavioral Classifications

BetterNight also looked at Dr. Mark Aloia’s 2008 research published in Annals of Behavioral Medicine. Dr. Aloia created 7 distinct phenotypes to describe CPAP usage patterns. These compliance categories ranged from highest compliance to lowest and included: good users, slow improvers, slow decliners, variable users, occasional attempters, early drop-outs and non-users. BetterNight looked at the behaviors that existed within each of these classifications when setting up the AI model and data sets. The goal was to move from reactive intervention to pro-active intervention. According to William Hevener, RGSPT, BetterNight’s Clinical Initiatives Manager, “I believe the earlier intervention will lead to higher adherence outcomes as well as increased staff efficiencies and the capability to manage larger populations for longer periods of time.”

BetterNIght’s study included a data set of 14,000 patients with a cross-sectional cohort of 3600 patients, CPAP usage data for 455 consecutive days, patient outreach notes and resupply data, and a yes or no classification for 90-day compliance. Further definition of the usage levels was established by ranking the previous 30 day period (most usage to least usage) and then by breaking down that group by specific time durations: greater than 4 hours of usage, less than 4 hours of usage and no usage. Ten 3-letter rating combinations were created and divided into 4 user phenotypes including good users, variable users, occasional users and non-users. This data set was then plugged into an AI model known as recurrent neural networks which uses the past to predict the future. The AI model produced a regression value prediction which forecasted how many hours next month patients will use CPAP and a classification of which phenotype the patient will belong to next month. The data go through simulations run millions of times, calculates errors and feeds them back into the model achieving breakthrough accuracy in forecasting. Further analysis led to over 20 options for behavioral outreach, trouble shooting and technical intervention.

The performance of the algorithm relative to CPAP usage was then evaluated using 3 different statistical analyses. The first analysis was for classification for the phenotype prediction. This was further broken down by reviewing 4 sub-analyses. The sensitivity analysis was done per phenotype and revealed how many patients the algorithm correctly identified as a specific user type. Specificity showed if the AI forecast was correct for the other groups not featured in the sensitivity analysis, accuracy combined sensitivity and specificity to give an overall sense of performance and Cohen’s Kappa took into account if the AI forecast was correct due to chance versus actual learning. The second statistical analysis was for the usage days prediction which measured the strength of the fit of the algorithm for matching the number of days the patients will use CPAP next month. The third statistical analysis was for resupply intervention. This analysis tested over 20 types of behavioral and technical intervention options and the average increase in monthly usage of the most effective interventions.

AI and Behavior Change: Future Directions in Patient Care

The statistical analysis of phenotype, usage prediction and resupply intervention revealed that CPAP usage varies month over month and that having personalized patient information helps to select the most meaningful intervention options. The Cohen’s Kappa score reflected the information garnered from AI was learned and not by chance. In the future, BetterNight will continue to refine and improve on these findings using AI. They will investigate using larger datasets of higher quality and improve on the rating and phenotype structure. Clinical validation for the AI results will be reviewed with more specific interventions being developed based on patient phenotype. Because patient behavior change is a critical driver for CPAP adherence, moving forward AI technology will further create and leverage tools that are at the intersection of this technology and behavioral change. Getting to the right patient at the right time is critical for successful intervention. BetterNight will continue to break new ground and set new standards of care by exploring the unique behaviors of each patient to best customize their care plan. BetterNight’s deep dedication to improving CPAP adherence by leveraging the most advance AI technology is further evidence that BetterNight is best positioned to deliver outstanding patient outcomes.

Learn more about the comprehensive BetterNight sleep apnea solution today.