Mood Prediction of Patients With Mood Disorders by Machine Learning Using Passive Digital Phenotypes Based on the Circadian Rhythm: Prospective Observational Cohort Study

Abstract

Background: Virtually, all organisms on Earth have their own circadian rhythm, and humans are no exception. Circadian rhythms are associated with various human states, especially mood disorders, and disturbance of the circadian rhythm is known to be very closely related. Attempts have also been made to derive clinical implications associated with mood disorders using the vast amounts of digital log that is acquired by digital technologies develop and using computational analysis techniques.

Objective: This study was conducted to evaluate the mood state or episode, activity, sleep, light exposure, and heart rate during a period of about 2 years by acquiring various digital log data through wearable devices and smartphone apps as well as conventional clinical assessments. We investigated a mood prediction algorithm developed with machine learning using passive data phenotypes based on circadian rhythms.

Methods: We performed a prospective observational cohort study on 55 patients with mood disorders (major depressive disorder [MDD] and bipolar disorder type 1 [BD I] and 2 [BD II]) for 2 years. A smartphone app for self-recording daily mood scores and detecting light exposure (using the installed sensor) were provided. From daily worn activity trackers, digital log data of activity, sleep, and heart rate were collected. Passive digital phenotypes were processed into 130 features based on circadian rhythms, and a mood prediction algorithm was developed by random forest.

Results: The mood state prediction accuracies for the next 3 days in all patients, MDD patients, BD I patients, and BD II patients were 65%, 65%, 64%, and 65% with 0.7, 0.69, 0.67, and 0.67 area under the curve (AUC) values, respectively. The accuracies of all patients for no episode (NE), depressive episode (DE), manic episode (ME), and hypomanic episode (HME) were 85.3%, 87%, 94%, and 91.2% with 0.87, 0.87, 0.958, and 0.912 AUC values, respectively. The prediction accuracy in BD II patients was distinctively balanced as high showing 82.6%, 74.4%, and 87.5% of accuracy (with generally good sensitivity and specificity) with 0.919, 0.868, and 0.949 AUC values for NE, DE, and HME, respectively.

Conclusions: On the basis of the theoretical basis of chronobiology, this study proposed a good model for future research by developing a mood prediction algorithm using machine learning by processing and reclassifying digital log data. In addition to academic value, it is expected that this study will be of practical help to improve the prognosis of patients with mood disorders by making it possible to apply actual clinical application owing to the rapid expansion of digital technology.

Keywords: circadian rhythm; digital phenotype; machine learning; mood disorder; projections and predictions; wearable device.

Conflict of interest statement

Conflicts of Interest: None declared.

©Chul-Hyun Cho, Taek Lee, Min-Gwan Kim, Hoh Peter In, Leen Kim, Heon-Jeong Lee. Originally published in the Journal of Medical Internet Research (http://www.jmir.org), 17.04.2019.

Figures

Figure 1

Comparison of basic feature distributions between high and low absolute mood score (HAMS and LAMS) groups. HAMS and LAMS were grouped into the upper or lower 10%, 30%, and 50% thresholds of the distribution. For each corresponding threshold group, we compared the main basic features of activity, light exposure, sleep, and heart rate (HR) related to the circadian rhythm (CR). The number in the parentheses on the horizontal axis means a cut-off threshold to distinguish HAMS and LAMS. (A) Steps_during_bedtime and (B) steps_during_daytime refer to activity levels of subjects during bedtime or daytime in each threshold HAMS or LAMS group. (C) Light_exposure_during_bedtime and (D) light_exposure_during_daytime refer to relative level of light exposure of subjects during bedtime or daytime. (E) Sleep_length and (F) sleep_efficiency refer to total sleep time (hours) and sleep efficiency (%), and (G) sleep_onset_dev and (H) sleep_offset_dev refer to regularity of sleep onset and offset time. (I) HR_CR_amplitude, (J) HR_CR_acrophase, and (K) HR_CR_mesor refer to the value of amplitude, acrophase, and mesor of cosine curve fitted HR, respectively. (L) resting_heart rate refers to the lowest HR at resting state during each day in the samples.

Figure 2

The performance evaluation of the mood state prediction model. The mood state prediction model outputs one of 2 mood states (ie, biased mood state or neutral mood state) and whether the model outcomes that were correctly matched with the ground truth (ie, the known actual mood states) was tested. The mood performance was evaluated in terms of the 4 performance evaluation metrics: sensitivity, specificity, accuracy, and area under the curve with the 3 different ground truth labeling criterion: 10%, 30%, and 50% cut-offs in absolute mood score distribution. (A) The performance evaluation result in the case of mood state labeling with 10% cut-off, (B) the performance evaluation result in the case of mood state labeling with 30% cut-off, and (C) the performance evaluation result in the case of mood state labeling with 50% cut-off. MDD: major depressive disorder; BD I: bipolar I disorder; BD II: bipolar II disorder; AUC: area under the curve.

Figure 3

Representative influential features of the mood state prediction model. The mood state prediction model uses several categories of features and different features have different contribution to prediction performance. Each of the feature categories was compared in terms of its importance (contribution perspective). Each bar in the graph means relative importance of a feature category compared to the other bars. The black sold lines at the end of each bar means one standard deviation error range, and the color coding of each bar means the direction of feature effect; therefore, sum of all the bar lengths is one (100%). (A) The comparison analysis was conducted with data from all patients. (B) The analysis was conducted only with data from major depressive disorder patients. (C) The analysis was conducted only with data from bipolar I disorder patients. (D) The analysis was conducted only with data from bipolar II disorder patients. MDD: major depressive disorder; BD I: bipolar I disorder; BD II: bipolar II disorder; HR: heart rate; CR: circadian rhythm.

Figure 4

Representative influential features of the mood episode prediction model. The mood episode prediction model uses several categories of features, and different features have different contributions to prediction performance. Each of the feature categories was compared in terms of its importance (contribution perspective). Each bar in the graph means relative importance of a feature category compared with the other bars. The black sold lines at the end of each bar means 1 SD error range, and the color coding of each bar means the direction of feature effect; therefore, the sum of all the bar lengths is one (100%). (A) The comparison analysis was conducted with data from all patients with depressive episodes. (B) The analysis was conducted only with data from patients with major depressive disorder with depressive episodes. (C) The analysis was conducted only with data from patients with bipolar type I disorder with depressive episodes. (D) The analysis was conducted only with data from patients with bipolar type II disorder with depressive episodes. (E) The analysis was conducted only with data from patients with bipolar type I disorder with manic episodes. (F) The analysis was conducted only with data from patients with bipolar type II disorder with hypomanic episodes. MDD: major depressive disorder; BD I: bipolar type I disorder; BD II: bipolar type II disorder; HR: heart rate; CR: circadian rhythm.