Among 199 individuals screened for participation, 22 were included in the GDM group and 15 in the non-GDM group (for study flowchart, see ESM Fig. 1). The age, education level, smoking status and mode of breastfeeding were not statistically different between the two groups (Table 1).

Table 1 Baseline clinical characteristics of the GDM and non-GDM groupsClinical measurements show increased metabolic dysfunction in the GDM group at 1–2 months postpartum

At baseline (1–2 months postpartum), the mean BMI was different between the GDM and non-GDM groups (mean difference +3.66 kg/m2 [95% CI 0.29, 7.03], p=0.03; Table 1); HbA1c also differed between groups (mean difference +2.85 mmol/mol [95% CI 0.35, 5.35 mmol/mol] or +0.26% [95% CI 0.03, 0.49%], p=0.03; Table 1 and ESM Fig. 2a). The difference in BMI and HbA1c was no longer significant in a sensitivity analysis without the five women with GDM and prediabetes at the first study visit. The other clinical measurements were not different between groups (Table 1).

Increased energy intake from fat and reduced energy intake from carbohydrates in the GDM group

Energy intake from carbohydrates was higher in the non-GDM group (mean ± SD 44.9±5.6%) than in the GDM group (mean ± SD 39.4±6.0%, p=0.02), while the energy intake from fat was lower in the non-GDM group (mean ± SD 39.0±4.2%) than in the GDM group (mean ± SD 43.5±5.9%, p=0.03; Table 2 and ESM Fig. 2b).

Table 2 Comparison of GDM and non-GDM groups across food and drink consumption measured with the smartphone app, CGM metrics and parameters extracted with the glucose model

The chrononutrition-related metrics derived from the MyFoodRepo app showed that the mean eating window and eating midpoint were similar between groups (Table 2). However, these values showed associations with glycaemic variables in a continuous analysis pooling data from both groups and time points (see below).

Higher glucose variability in the GDM group, even in the absence of prediabetes

While the mean glucose and CV calculated from the CGM data were not different between groups at baseline (both p≥0.14; Table 2 and ESM Fig. 2c), MAGE was higher in the GDM group (median 1.46 [IQR 1.32–1.69] mmol/l) than in the non-GDM group (median 1.26 [IQR 1.08–1.38] mmol/l, p=0.04). Furthermore, when women with prediabetes at baseline were excluded in a sensitivity analysis (5/22 in the GDM group), the difference in MAGE persisted (p=0.02).

Comparing the mean 24 h glucose profiles between groups did not reveal a clear signal due to high interindividual heterogeneity (ESM Fig. 3). This motivated the use of a computational model combining ingestion events and CGM data to better characterise the glucose dynamics for each participant.

Glucose decrease after meal consumption is slower in the GDM group

Figure 1a shows a schematic of the glucose dynamics model, which estimates the height of the glucose spikes (termed ‘response heights’) after food or drink consumption, as well as the time required for glucose to return to baseline after a spike (termed ‘response t½’). Among these glucose model parameters (Table 2), the response t½ was ~30 min longer (i.e. slower postprandial decrease of glucose) in the GDM group than in the non-GDM group at baseline (p=0.01, Fig. 1b); this was in the absence of higher carbohydrate intake in women with GDM (Table 2). As an example, the participant with the shortest response t½ (participant ID 03, who was in the non-GDM group) showed rapid responses to meals, with glucose returning quickly to baseline levels after meals (Fig. 1c). In contrast, the participant with the longest t½ (ID 29, who was in the GDM group) showed comparatively much slower glucose responses, where glucose excursions visibly took much longer to return to baseline levels after ingestion events (Fig. 1d).

Lower amplitude of the underlying glucose 24 h rhythm in the GDM group

The glucose dynamics model also provides the underlying glucose 24 h rhythm (schematic shown in Fig. 2a), with the amplitude being higher in the non-GDM group than in the GDM group at baseline (i.e. with more pronounced oscillations around the 24 h clock [p=0.005, Table 2, Fig. 2b]). The participant with the highest amplitude showed a regular underlying 24 h rhythm, with ingestion events causing further glucose spikes on top of the oscillating trendline (ID 05, who was in the non-GDM group, Fig. 2c), while this 24 h rhythm was barely detectable in the participant with the lowest amplitude (ID 17, who was in the GDM group, Fig. 2d). The amplitude of the underlying glucose 24 h rhythm did not affect overall glucose variability. (see ESM Results and ESM Fig. 4 for further details).

Fasting glucose and cholesterol evolve during the postpartum period

We assessed the evolution of clinical and glycaemic variables from baseline (1–2 months postpartum; median 42 [range 26–65] days since delivery) to follow-up (6 months later; median 221 [range 200–285] days since delivery). At follow-up, there was no difference in breastfeeding mode between groups (p=0.45). The comparison across both groups revealed a decrease in waist circumference, total cholesterol and LDL-cholesterol (all p≤0.001) and an increase in fasting plasma glucose (p=0.02) over time (ESM Fig. 5a), but no between-group differences in these changes (ESM Fig. 5b) (see ESM Results for further details).

Glycaemic variables are correlated with sleep behaviour and ‘eveningness’ across both groups

Pooling baseline and follow-up data from both groups to increase power in a linear mixed model confirmed the known association between BMI and both HbA1c and fasting plasma glucose (Fig. 3a, b). We also found that a high amplitude of glucose 24 h rhythm was associated with more physical activity (as measured with the International Physical Activity Questionnaire, p=0.01, Fig. 3c), high HDL-cholesterol and low triacylglycerols (both p≤0.004, ESM Fig. 6).

Fig. 3

In a linear mixed model pooling all baseline and follow-up data, glycaemic variables are associated with clinical measurements and lifestyle factors as measured with questionnaires. (a) Points show the inferred regression coefficient, horizontal lines the 95% CI; p values are also shown. The dependent variables were log-transformed (natural log). (b–e) Examples of regressions for specific variables. β denotes the inferred regression coefficient (with 95% CI), which is also shown graphically (grey line and shaded blue, respectively). Blue symbols, non-GDM group; red symbols, GDM group; circles, baseline; triangles, follow-up. MET, metabolic equivalent of task; quest, questionnaire

We then explored the relationships between glycaemic variables and sleep behaviour and chronotype as assessed with questionnaires. Late sleep midpoint was associated with higher fasting plasma glucose and weaker amplitudes in the underlying 24 h glucose rhythm (both p=0.02, Fig. 3d, e). The eveningness score and a short sleep duration were associated with higher fasting plasma glucose (p=0.049 and 0.03, respectively, Fig. 3a).

Wearable device-derived markers of late circadian timing are associated with higher fasting plasma glucose and lower amplitude of glucose 24 h rhythm

Finally, we again pooled baseline and follow-up data within a linear mixed model framework to quantify the relationships between circadian timing metrics derived from the MyFoodRepo, GENEActiv and Actiheart wearables and glycaemic variables (Fig. 4a).

Fig. 4

The relationship between wearable device-derived timing metrics and glycaemic variables. (a) The β regression coefficients between glycaemic and timing variables are shown as a heat map. The dependent variables were log-transformed (natural log). *p<0.05, **p<0.01. (b–e) Examples of regressions for specific variables. β denotes the inferred regression coefficient (with 95% CI), which is also shown graphically (grey line and shaded blue, respectively). Blue symbols, non-GDM group; red symbols, GDM group; circles, baseline, triangles, follow-up. HR, heart rate; HRV, heart rate variability

The eating midpoint (i.e. the daily midpoint of all food and drink consumption) and eating duration were similar between non-GDM and GDM groups at baseline (Table 2). However, regression analysis across both time points revealed that an early eating midpoint and shorter eating duration were correlated with a higher amplitude glucose 24 h rhythm (p=0.048 and p=0.008, respectively, Fig. 4b, c).

A similar analysis of wearable device-derived sleep metrics showed that a late sleep midpoint was associated with a lower amplitude glucose 24 h rhythm (Fig. 4d), and shorter sleep was associated with a longer glucose response t½ (Fig. 4e). Among the heart activity timing metrics, the strongest association was between a later peak time of the heart rate and higher fasting plasma glucose (Fig. 4a).