Intelligent phenotyping with machine learning for therapeutic personalization in a digital sleep improvement program

Abstract

Sleep disturbances are prevalent, associated with relevant health outcomes, and generally underdiagnosed. More than 80 disorders are described by the International Classification of Sleep Disorders (AASM, 2014), being insomnia and sleep-disordered breathing the most frequent. The overlap between these two conditions is also high, with ~50% of patients with sleep apnea also experiencing insomnia. Data from the EPISONO study show that about 1/3 of the adult population report chronic insomnia, while approximately 1/10 have severe apnea, and both conditions double the risk for emergency care and hospitalizations in two years.Vigilantes do Sono's mission is to help people to identify and care for their sleep problems, with the vision of becoming a public health strategy. Its current product is a digital Cognitive-Behavioral Therapy for Insomnia (CBT-I). It is innovative and the first built with chatbot and artificial intelligence (AI). It includes ~51 short sessions (5-10 minutes), distributed in seven modules, and guided by a virtual assistant. So far, more than 30,000 people have accessed the application, around 600 health professionals have registered on the telemonitoring platform, and several companies have hired the solution as a benefit for employees/customers. The program's feasibility study indicated its accessibility, usability and effectiveness. About 86% of users who completed all modules reported improvements on sleep quality, with an average increase of 67.3 (52.8-81.8) minutes in total sleep time. The benefits and cost-effectiveness of the digital program are clear, in addition to its potential social impact, due to the reduction in healthcare costs and increased productivity. However, the proportion of individuals with insomnia who do not engage in or respond to therapeutic modalities is large. Different reasons can explain these effects. It is reasonable to assume that the overlap with other sleep disorders, such as sleep apnea, is one of them. In addition, one of the main complaints related to disengagement is the need to fill in sleep diaries, which are crucial for personalizing therapy. Currently, the program relies on users manually entering sleep data. In this sense, modules allowing for the acquisition of objective sleep related data can increase the program's accessibility, engagement and effectiveness.Thus, the present study aims to develop digital phenotyping methods by capturing the smartphone's accelerometer, microphone, and camera data, to measure the sleep pattern and anthropometric, craniofacial and intraoral characteristics related to the risk of sleep apnea, as well as by integrating objective and subjective data into the AI, enhancing machine learning.Specific questions involving risk for sleep apnea and other comorbidities will be added along with those already available in the app. Modules for the assessment of objective sleep data, the occurrence of snoring, and measurements of neck circumference and craniofacial images will be developed and modeled with volunteers using the application. They will be recruited and randomly referred to polysomnography either in the laboratory or at home, and simultaneously, in both groups, they will be evaluated by the tools incorporated into the application. All volunteers will use actigraphs for 14 consecutive days and, on each night, they will also be simultaneously monitored by the application with the smartphone.The application of supervised machine learning integrating subjective and objective sleep data, to indicate overlapping between sleep disorders and other clinical conditions, will improve the phenotype identification, enriching customization and content adaptation by the AI, expanding the scope and cost-utility of services offered by Sleep Watchers. (AU)