Digital circadian and sleep health in individual hospital shift workers: A cross sectional telemonitoring study

Abstract

Background: Telemonitoring of circadian and sleep cycles could identify shift workers at increased risk of poor health, including cancer and cardiovascular diseases, thus supporting personalized prevention.

Methods: The Circadiem cross-sectional study aimed at determining early warning signals of risk of health alteration in hospital nightshifters (NS) versus dayshifters (DS, alternating morning and afternoon shifts). Circadian rhythmicity in activity, sleep, and temperature was telemonitored on work and free days for one week. Participants wore a bluetooth low energy thoracic accelerometry and temperature sensor that was wirelessly connected to a GPRS gateway and a health data hub server. Hidden Markov modelling of activity quantified Rhythm Index, rest quality (probability, p1-1, of remaining at rest), and rest duration. Spectral analyses determined periods in body surface temperature and accelerometry. Parameters were compared and predictors of circadian and sleep disruption were identified by multivariate analyses using information criteria-based model selection. Clusters of individual shift work response profiles were recognized.

Findings: Of 140 per-protocol participants (133 females), there were 63 NS and 77 DS. Both groups had similar median rest amount, yet NS had significantly worse median rest-activity Rhythm Index (0·38 [IQR, 0·29-0·47] vs. 0·69 [0·60-0·77], p<0·0001) and rest quality p1-1 (0·94 [0·94-0·95] vs 0·96 [0·94-0·97], p<0·0001) over the whole study week. Only 48% of the NS displayed a circadian period in temperature, as compared to 70% of the DS (p=0·026). Poor p1-1 was associated with nightshift work on both work (p<0·0001) and free days (p=0·0098). The number of years of past night work exposure predicted poor rest-activity Rhythm Index jointly with shift type, age and chronotype on workdays (p= 0·0074), and singly on free days (p=0·0005).

Interpretation: A dedicated analysis toolbox of streamed data from a wearable device identified circadian and sleep rhythm markers, that constitute surrogate candidate endpoints of poor health risk in shift-workers.

Funding: French Agency for Food, Environmental and Occupational Health & Safety (EST-2014/1/064), University of Warwick, Medical Research Council (United Kingdom, MR/M013170), Cancer Research UK(C53561/A19933).

Keywords: Body temperature; Circadian rhythm; Rest-activity; Shift work; Sleep; Telemonitoring.

Figure 1

a: Consort diagram. Consort diagram showing the enrolment of subjects in the study and selection of subjects whose records that were adequate on both work and free days for analysis. Subjects are stratified according to shift type (day shift or night shift). No participant alternated day and night shift. *Heat-related disconnection of electronic cards within the GPRS gateway. b: Analysis Chart. The template for the statistical analysis of multidimensional data sets including physical activity and skin temperature. ‘EDA’ denotes the exploratory data analysis.

Figure 2

a: State estimation and spectral analysis for subject 1219 (39 y.o. female, DS). b: State estimation and spectral analysis for subject 1363 (45 y.o. female, NS). (i) Time series of PA (black dots) and Chesttemp (brown dots) with yellow line indicating the most likely state using local decoding. The red and blue horizontal bars denote the work and sleep periods recorded in the diary. This plot shows that the rest state dominated the sleep period recorded in diary and that there were a lot of transitions between the MA and HA state during the remaining monitoring period. (ii) Spectral density estimates of Chesttemp (brown line) with respective 90% confidence intervals (grey area). (iii) State probability plot during the whole study period, i.e. cumulative plot of $\text{Pr}(S_t=j|O^{\left(T\right)})$for $j$ = 1 (IA, blue), 2 (MA, pink), 3 (HA, red). It allows a quick assessment of how probable the most likely state is and what other states have noticeable probability and provides visual information on how well a subject has rested. In particular, if they have solid blue areas, i.e. rarely move into the active states during rest, then it could be summarized that this subject has obtained a good rest, as the subject 1219 has done. (iv) Circadian state probability plot from HHMM obtained as the averaged probability of every state over the whole study period. It shows the periodic time profile of the three state probabilities plotted in cumulative manner analogous to (iii) with same colour coding as in (iii). The blue area represents the probability of rest and gives the daily profile of rest. Subject 1219 had one rest every day shown by a(iv) while subject 1363 was likely to rest twice in one day illustrated by two peaks in the rest profile in b(iv).

Figure 3

Daily profile of rest for two shift groups and boxplot of individual circadian parameters. (a-b) Plots of probability of rest for all 77 DS (yellow dashed lines) with group median (black dashed line) in left panel and for all 63 NS (blue dashed lines) with group median (black dashed line) in right panel. The individual probability of rest line is the same as the boundary of blue part in Figure 2a(iv) and 2b(iv). The group median is obtained by taking the median of probabilities of rest for all members at each time point. (c-f) Plots of the four circadian parameters during the whole study period stratified by shift type. Boxes with yellow dots are for DS, boxes with blue dots are for NS. The median of individual p1-1 for DS is 0·96 [IQR, 0·94-0·97] while for NS it is 0·94 [0·94-0·95]. The median of individual RI for DS is 0·69 [0·60-0·77] while for NS it is 0·38 [0·29-0·47]. The median of individual rest amount for DS is 8·9 [7·9-9·5] while for NS it is 8·2 [7·3-9·3]. The median of individual centre time of rest for DS is 3·3 [2·8-3·9] while for NS it is 6·9 [6·0-8·9]. (g-j) Plots of the four circadian parameters on workdays and free days separately stratified by shift type. Yellow boxes are for DS, blue boxes are for NS. Within each shift type, boxes with black dots (left) represent free days, boxes with red dots represent workdays (right). Take (g) as an example. Comparing the two boxes with black (red) dots is to compare the p1-1 on free (work) days between DS and NS. Comparing the left (right) two boxes is to compare the p1-1 between work and free days for DS (NS). In (j), the original value of centre time denoted by ‘Ori’ is the gravity centre of the area under the probability of rest curve, while the new value (‘New’) is the timing of the maximum probability of rest between 07:00 and 20:00. Numerical representation of boxplots can be found in Table S2A-B.

Figure 4

Spectral density estimation of Chesttemp for two shift groups. (a-b) Plots of spectral density estimates (black dashed line) for every subject with group median (blue/yellow dashed line) and 10%-90% quantile (blue/yellow area). The individual spectral density estimate line is the same as Figure 2a(ii) and 2b(ii). The group median is obtained by taking the median of the densities for all DS/NS. The lower and higher boundary of the light blue/yellow ‘envelope’ are equal to the 10^th and 90^th percentiles across the individual densities, respectively. These ‘envelopes’ provide an idea of where most individual densities are. Besides the dominant 24h rhythm, a dominant circa-12h rhythm, with a period ranging from 10h to 14h, was identified in 12 DS (16%) and 20 NS (30%). There were 2 DS (3%) and 6 NS (10%), who displayed a dominant period around 8h (7h-9h), and a further 9 DS (12%) and 8 NS (13%) with other dominant or undetectable periods. (C) Boxplots of the spectral gravity centres across all individuals. The median gravity centre for DS is 15·6h [IQR, 14·0 to 18·0] and for NS is 14·3h [12·5 to 16·8]. (d-f) Magnifications of the plots (a-b) showing group median estimates (yellow for DS, blue for NS) and 10%-90% quantile (yellow for DS, blue for NS) in three intervals (2-9h, 9-16h, 16-40h). A larger proportion of DS had a 24h dominant period, resulting in a sharp peak of the group median line and the group envelope in (a) and (f). The flatter and wider peak at 24h for NS group illustrates more variability in the dominant periods around 24h in (b) and (f). The magnified plots in (d) highlight the fact that NS had more pronounced spectral peaks at shorter periods, especially around 8h.

Figure 5

Three DS clusters and three NS clusters. (a) (b) (e) (f) (i) (j) Plots of probability of rest for all DS (cluster 1: red dashed lines; cluster 2: yellow dashed lines; cluster 3: blue dashed lines) with cluster median (black dashed lines), similar to Figure 3a. The cluster median is obtained by taking the median of probabilities of rest for all members at every time point. There are 34 (44%), 24 (31%), and 19 (25%) subjects in cluster 1, 2 and 3 respectively. (c) (d) (g) (h) (k) (l) Plots of probability of rest for all NS (cluster 1: purple dashed lines; cluster 2: orange dashed lines; cluster 3: green dashed lines) with cluster median (black dashed lines) similar to Figure 3b. The cluster median is obtained by taking the median of probabilities of rest for all members at every time point. There are 28 (44%), 11 (17%), and 24 (38%) subjects in cluster 1, 2 and 3 respectively. (m) (o) Plot of cluster median obtained by taking the median of the densities for all DS/NS subjects in one cluster. (n) (p) Histogram of dominant period in Chesttemp for three DS/NS clusters. (q) Boxplot of individual sub-periods for every cluster. The individual sub-period is the median of sub-periods over several rest bouts. The labels underneath the x_axis denote the Spearman correlation coefficient between sub-periods of LIDS PA and Chesttemp.