JMIR Publications

ABSTRACT

Background:

In the modern economy, shift work is prevalent in numerous occupations. However, shift work often conflicts with the workers’ circadian rhythm and can result in shift work sleep disorder (SWSD). Proper management of SWSD emphasizes comprehensive and patient-specific strategies and some of these strategies are analogous to the cognitive behavioral treatment of insomnia (CBTI).

Objective:

In this paper, we aim to develop and evaluate machine learning algorithms that predict physicians’ sleep advice using wearable and survey data. We developed an online system to conveniently and frequently provide individualized sleep and behavior advice with CBTI elements for shift workers.

Methods:

Data were collected for a period of 5 weeks from shift workers in the ICU at two hospitals (N = 61) in Japan. The data were composed of three modalities, (1) Fitbit data, (2) survey data, and (3) sleep advice. We handcrafted physiological and behavioral features from the raw data and identified clusters of participants with similar characteristics using hierarchical clustering. After the first week of enrollment, physicians reviewed Fitbit and survey data to provide sleep advice from a list of 23 messages. We implemented random forest (RF) models to predict the 7 most frequent messages given by the physicians. We tested our predictions under participant dependent and independent settings and analyzed the most important features for prediction.

Results:

We found that the clusters were distinguished by work shifts and behavioral patterns. For some clusters, having a work shift on a given day contributed to low wellbeing scores on that day. Another cluster had days with low sleep duration and the lowest sleep quality when there was a day shift on the day before and a midnight shift on the current day. Our advice prediction models achieved higher F1 scores in 27 of 28 t-tests conducted, and the performance differences were statistically significant with P < .001 for 24 tests and P < .05 for 3 tests compared to the baseline. The analysis of the feature importance of our models showed that the most important features matched the message sent to participants. For instance, for message 7 (darken the bedroom when you go to bed), the models primarily examined the average brightness of the sleep environment to make predictions.

Conclusions:

Although our current system requires physician input, an accurate machine learning algorithm would be promising for automating without hurting the trustworthiness of selected recommendations. The algorithm is limited to the 7 most popular ones among 23 choices due to rare occurrences of the remaining options. Therefore, further studies are necessary to gather enough data to enable predictions for less frequent advice labels. Clinical Trial: UMIN Clinical Trials Registry UMIN000036122 (phase 1), UMIN000040547 (phase 2); https://tinyurl.com/dkfmmmje, https://upload.umin.ac.jp/cgi-open-bin/ctr_e/ctr_view.cgi?recptno=R000046284.