• korea
  • search
  • login


AI Algorithm for Predicting Sleep Quality

2020-10-05 11:04:53

Professor MeeYoung Cha's research team in KAIST School of Computing developed an interpretable deep learning algorithm that can predict the sleep quality of insomnia patients as well as the priority ranking of insomnia symptoms among patients based on wearable device data.

Sleep quality is an important factor that determines a person's condition of the day. Insomnia is a serious social problem; it is declared as an epidemic of public health by the U.S. Centers for Disease Control and Prevention. It can lead to not only various diseases but also loss of labor at the national level. Precision Psychiatry is a recently established research field on detailed modeling of various diseases based on Big Data. Like precision medicine, it brings innovation to provide personalized services based on various characteristics of individuals when proposing interventions or therapy for mental disorders.

KAIST Data Science Lab (Advisor: Professor MeeYoung Cha) conducted a study to predict sleep quality by collecting and analyzing sleep patterns and individual behavioral data from wearable devices at a perspective of Big Data. This study was conducted in collaboration with researchers in Computer Science and Mental Health, and the results of the study were published at ACM SIGKDD Conference on Knowledge Discovery and Data Mining (Title: Learning Sleep Quality from Daily Logs, KDD 2019) and JMIR mHealth and uHealth (Title: Clustering Insomnia Patterns by Data from Wearable Devices: Algorithm Development and Validation, JMU 2019, Impact Factor 4.541). ACM SIGKDD is the top conference in the field of Data Mining, and JMIR mHealth and uHealth is a renowned journal in the field of digital health.

First, in this study, the researchers proposed an interpretable deep learning model that predicts the sleep quality of insomnia patients based on wearable device data. Sleep quality can be measured in various ways, among which the researchers used sleep efficiency (actual sleep time/total time in bed). They collected behavioral and sleep data from 42 subjects during consecutive 24 hours for 6 weeks, converted the data into vectors for each day, and then continuously injected the vectors into the LSTM-Attention ensemble deep learning model (Figure 1.) to predict the daily sleep efficiency index. Using this model, the prediction error rate was reduced by 9-16% compared to the existing methods, and it was possible to explain factors that would influence future sleep efficiency for each user.

Next, this study predicted the risk ranking for each patient group from the perspective of medical experts. Using this information, it is possible to preemptively inform the doctors which of the insomnia patients need more care and attention. These priorities help the limited resources of the medical staff lead to more effective online counseling. To this end, the researchers developed a model for continuous feature learning with the premise that users with a high or low risk of potential insomnia would have similar behaviors and sleep patterns. The model compares the daily characteristics of each user to extract the phenotype, which is injected into the deep learning ranking prediction model to predict the final sleep efficiency ranking among users. The proposed model was able to predict the daily ranking relationship more accurately compared to the existing methodology. Moreover, it was able to learn the interrelationships between features at the same time.

Professor Meeyoung Cha, the head of this research, said, “After the coronavirus outbreak, the biorhythm of the whole population broke down, and there are many adolescents and adults suffering from sleep problems. The AI ​​model and sleep intervention program are still in infancy, but in the next study, we hope to develop this model into an app that can be tested and used for a wider range of experiments."

< Figure 1. LSTM-Attention ensemble deep learning model >

< Figure 2. Observed and predicted sleep efficiency (left), Interpretation of sleep efficiency prediction results (right) >

< Figure 3. The basic idea of PLRG (left), Prediction scores of priority ranking among users with insomnia (right) >