Targeting Habitual Fast Eating as a Modifiable Risk Factor for Type 2 Diabetes: A Propensity Score-Matched Analysis

Abstract

Introduction

Type 2 diabetes remains a growing public health concern, and modifiable lifestyle behaviors represent a key target for prevention. Eating habits have been proposed to influence metabolic risk, yet evidence from large real-world cohorts with rigorous covariate control is limited. We hypothesized that habitual eating habits, especially eating speed, are a modifiable risk factor for incident type 2 diabetes in a Japanese adult population.

Methods

This retrospective observational study used lifestyle habit data from medical check-up records collected in Shizuoka, Japan (April 2008–March 2020). After applying exclusion criteria, 98,718 subjects aged 18–75 years were eligible for analysis. The analytic sample for each eating habit varied after propensity score matching. Cox proportional hazard regression was used to assess new-onset type 2 diabetes in relation to multiple eating behaviors, including eating speed, breakfast skipping, late dinners, post-dinner snacking and frequency of vegetable consumption. Covariate adjustment was performed using propensity score matching for sex, age, BMI, smoking and drinking status, blood pressure and LDL and HDL cholesterol levels.

Results

During the observation period, 4631 new cases of type 2 diabetes were identified. After full adjustment, fast eating was significantly associated with increased diabetes risk (HR: 1.12; 95% CI: 1.02–1.23; p = 0.01). Sex-stratified analysis showed that fast eating among women was associated with a 14% higher risk (HR: 1.14; 95% CI: 1.00–1.30; p = 0.03), while no statistically significant association was observed in men.

Conclusion

Habitual fast eating may represent a modifiable risk factor for type 2 diabetes, particularly among women, though the association was modest and warrants confirmation in future clinical studies.

Supplementary Information

The online version contains supplementary material available at https://doi.org/10.1007/s13300-026-01912-1.

Key Summary Points

Why carry out this study?

Type 2 diabetes poses a substantial and growing burden in Japan, yet modifiable dietary behaviors such as eating speed remain insufficiently studied as targets for prevention in large real-world cohorts.

This study investigated whether habitual fast eating and other eating behaviors are associated with incident type 2 diabetes in Japanese adults, using propensity score matching to minimize confounding.

What was learned from the study?

Habitual fast eating was significantly associated with a 12% increased risk of developing type 2 diabetes after full covariate adjustment, with a more pronounced 14% elevated risk observed specifically among women.

These findings identify habitual fast eating as a modifiable behavioral target that may be incorporated into lifestyle-based diabetes prevention strategies in clinical practice and public health interventions.

Introduction

Diabetes mellitus has become a major global health challenge in the twenty-first century. It imposes significant burdens on individuals, healthcare systems and economies. The Global Burden of Disease Study 2021 showed diabetes to be the eighth leading cause worldwide of death and disability, affecting an estimated 529 million people [1]. More than 1.31 billion people are projected to be living with diabetes by 2050 [1]. Most of this burden derived from type 2 diabetes might be largely preventable and even, in some cases, reversible by managing key risk factors [1].

Japan has not been immune to this worsening global trend. The nation has experienced a marked increase in the prevalence of diabetes or impaired glucose tolerance over the past few decades [2, 3]. Japan is moving toward having the highest number of adults with diabetes among the top 10 countries. Thus, there is an urgent need for effective population-specific prevention strategies [4].

Among several modifiable lifestyle factors, eating behaviors have attracted substantial scientific interest as key determinants of metabolic health. While dietary composition is well known to be important, emerging evidence also indicates that the timing and manner of food intake play equally crucial roles in the pathogenesis of type 2 diabetes. These behaviors exert direct influences on metabolic processes, impacting glucose tolerance, insulin sensitivity and weight management, which play central roles in the development of diabetes [5]. Therefore, research has increasingly focused on the influence of eating behaviors. A large-scale study conducted on the Japanese population identified fast eating as a strong and independent risk factor for new-onset diabetes, even after adjusting for confounding variables including body mass index (BMI) and alcohol consumption [6]. Other studies have confirmed this finding, suggesting eating speed to be a novel and important target for intervention [7, 8].

While the manner in which meals are consumed matters, when they are eaten also has relevance. Chrononutrition studies suggest that meal timing can significantly impact glucose tolerance and diabetes risk by influencing the body’s circadian rhythms [9]. Habits such as eating late dinners or snacking after the evening meal have been linked to impaired glucose control and a higher risk for type 2 diabetes [10, 11].

Beyond eating timing and speed, lack of meal regularity, especially skipping breakfast, has consistently been identified as an independent risk factor for developing type 2 diabetes. Multiple large-scale meta-analyses from global cohort studies have shown a positive link between skipping breakfast and a higher diabetes risk [12, 13]. The underlying physiological mechanisms involve the harm to metabolic processes of missing the first meal of the day. Specifically, skipping breakfast was found to be associated with increased postprandial glucose spikes and to reduce insulin sensitivity at later meals, i.e., the notion of impaired metabolism manifesting as the “second meal effect” [14]. Regular breakfast consumption is now widely recognized as a simple and powerful factor in preventing metabolic disease.

The link between individual lifestyle factors and diabetes has increasingly been clarified. There is a need for large-scale, long-term observational studies examining these habits collectively within a specific, well-defined population. A significant challenge in observational research is the potential for confounding variables to bias results. To address this possibility, modern statistical techniques, such as propensity score matching (PSM), can simulate many features of randomized controlled trial conditions and thereby strengthen causal inferences from non-intervention or non-prospective cohort databases [15, 16].

This study was designed to examine how a group of lifestyle habits, including skipping breakfast, fast eating, post-dinner snacking, late meals and eating vegetables at every meal, impact the development of diabetes. We studied 98,718 participants from Shizuoka prefecture, Japan, using medical check-up data collected from April 1, 2008, starting with an identical lifestyle questionnaire in Japan, until March 31, 2020, just before the COVID-19 pandemic. Using propensity score analysis, we evaluated how these common behaviors influence diabetes risk in the Japanese population in hopes of obtaining valuable insights applicable to public health policy and clinical guidance.

Methods

Ethical Approval

This study was conducted according to the guidelines laid down in the Declaration of Helsinki, and all procedures involving research study participants were approved by the University of Shizuoka Ethics Committee (approval no. 6–3). After approval, consent was obtained using an opt-out form posted at the medical institution and on its website. Data were anonymized by the medical institution before transfer to the research team, and the data usage agreement between the medical institution and the research team specified that the data would be used for research purposes only.

Study Design and Subjects

This was a retrospective observational cohort study comprising all individuals who underwent annual medical check-ups provided by the Seirei Social Welfare Corporation. The analysis used a dataset containing comprehensive health information of pseudonymized medical check-up records provided as Comma-Separated Values formatted text files over a multi-year period (2008 to 2020). The flow diagram showing the exclusion and inclusion criteria is presented as Supplementary Fig. 1. From a possible total of 166,968 participants, 57,304 subjects with only one health check-up, 3766 diagnosed with diabetes at the 1 st or 2nd consecutive year of the study, 565 with a history of cancer, 645 with pancreatic disease, 3445 with liver disease and 2525 with kidney disease were excluded. To minimize confounding bias arising from temporal changes in dietary behavior, screening was performed to identify only subjects with consistent self-reporting of eating habits, so-called habitual behaviors, from the initial check-up and extending throughout the study period. After exclusions, 98,718 subjects [55,507 men (56.2%) and 43,211 women (43.8%)] remained eligible for analysis, with the analytic sample for each eating habit varying after propensity score matching (e.g., n = 50.180 for fast eating; see Supplementary Table 1 for all variables). For each lifestyle variable, only participants who maintained a consistent answer, whether consistently yes or consistently no, or consistently fast or normal/slow, throughout the entire study period were included in the analysis of that specific variable. Participants who changed their response for a given eating habit at any point during follow-up were excluded from the analysis of that variable only and were still eligible for analyses of other variables for which their answers had remained stable. This approach was applied independently for each variable, meaning that the analytical sample size varies per eating habit but in no instance excludes a particular response category. Participants receiving lipid-lowering medications (e.g., statins) at baseline were not excluded from the analysis; lipid-related parameters were instead accounted for through covariate adjustment and propensity score matching, as detailed below.

Assessment of Eating Habits

The primary exposures of interest were five specific eating habits, which were ascertained from self-reported questionnaires administered during the annual health check-ups. Skipping breakfast was defined as three or more times a week. Fast eating was defined based on the participant’s self-perception of their eating speed, as compared with others, as “slow,” “normal” or “fast.” Snacking after dinner was defined as the regular consumption of snacks after the evening meal. Dinner before bedtime was characterized by the habitual consumption of the evening meal within ≤ 2 h before bedtime. Eating vegetables at every meal was defined as the habit of eating at least one portion of vegetables daily at every meal consumed.

Study Endpoint

The primary endpoint of the study was the incidence of new-onset type 2 diabetes during the follow-up period. Diabetes was defined as fasting plasma glucose ≥ 7.0 mmol/l (126 mg/dl) and/or HbA1c ≥ 6.5%. Also, diabetes was confirmed based on a diagnosis history from a medical clinic or hospital, or medication prescribed before the next medical check-up. Starting in 2013, the method of calculating HbA1c changed from the Japanese Diabetes Society (JDS) to the National Glycohemoglobin Standardization Program (NGSP) criteria. The HbA1c values from 2008 to 2012 were therefore adjusted accordingly [17].

Statistical Analysis

Proportional hazard assumptions were assessed using Schoenfeld residuals and the complementary loglog method. To minimize the confounding bias inherent to observational studies and to simulate the characteristics of a randomized controlled trial, we applied the PSM method with adjustments for sex, age, BMI, blood pressure, LDL and HDL cholesterol, smoking and drinking status. Propensity scores were estimated using logistic regression models for each eating habit, based on the covariates described above (sex, age, BMI, blood pressure, LDL and HDL cholesterol, smoking and drinking status). Matching was performed using the MatchIt package (version 4.7.2) in R, using nearest neighbor matching with a 1:1 ratio and a caliper width of 0.2 standard deviations of the logit of the propensity score, without replacement. Covariate balance after matching was assessed using standardized mean differences (SMD), with SMD < 0.1 considered indicative of adequate balance (Table 1). For the sex-stratified analyses, matching was performed independently within each sex stratum using the same algorithm, ratio and caliper width as the overall analysis. Because caliper matching does not guarantee a match for every subject, and the pool of eligible controls differed between strata, the sum of matched subjects across the male and female strata is smaller than the total matched sample size (for eating speed) from the pooled analysis. The adjusted covariates were selected based on past similar analyses, including those using the PSM method [16, 18, 19]. The associations between eating behaviors and outcomes were evaluated using a Cox proportional hazards model after PSM. Each eating habit was analyzed in a separate multivariable Cox proportional hazards model rather than in a single model mutually adjusted for all five habits. This approach was chosen for two reasons. First, because participants were required to have consistent responses only for the specific habit under analysis (Section “Study Design and Subjects”), requiring simultaneous consistency across all five variables would substantially reduce the eligible sample size for each model. Second, several of these eating behaviors are behaviorally correlated (e.g., skipping breakfast and snacking after dinner may co-occur as part of a broader irregular eating pattern), raising the potential for multicollinearity if entered into the same model. This approach is consistent with prior studies examining multiple eating habits in relation to diabetes risk in Japanese cohorts [20]. For the variable “fast eating,” the categories “slow” and “normal” were combined for the analysis, since Cox regression can only be applied to analyzing binary values. As the “skipping breakfast” variable violated the proportional hazard assumptions when analyzing the entire study population, time-restricted Cox regression was performed. Parametric continuous variables are presented as means with standard deviation. The categorical variables are presented as absolute numbers and percentages. All statistical analyses were performed using the latest version of RStudio, and differences were considered statistically significant at p ≤ 0.05.

Table 1.

Baseline characteristics of subjects analyzed in this study

Variable	Before matching	After matching					SMD values per variable after matching
N	98.718	See Supplementary Table 1
		BR	DIN	FAST	SNK	VEG	BR	DIN	FAST	SNK	VEG
Male (%)	55.507 (56.2)	9.436	17.718	12.998	7.262	13.012	< 0.01	0.02	< 0.01	0.01	0.01
Female (%)	43.211 (43.8)	3.628	5.584	7.168	7.784	8.120
Age, years	48.0 ± 10.7	47.0 ± 5.4	46.0 ± 5.8	48.0 ± 5.1	48.0 ± 5.2	49.0 ± 5.0	0.02	0.04	< 0.01	0.02	< 0.01
BMI, kg/m2	22.5 ± 3.2	22.5 ± 3.0	22.3 ± 3.2	22.4 ± 3.1	22.5 ± 3.2	22.4 ± 3.0	< 0.01	< 0.01	< 0.01	< 0.01	< 0.01
SBP, mmHg	119 (109–129)	118 (109–129)	119 (109–129)	119 (108–127)	119 (108–127)	118 (109–129)	0.01	0.03	0.02	0.02	0.02
DBP, mmHg	76 (66–85)	76 (66–85)	76 (66–81)	76 (66–85)	76 (68–85)	76 (66–85)	0.03	0.01	0.02	0.03	0.03
LDL-C, mg/dl	117 (102–137)	117 (102–137)	119 (105–137)	116 (102–135)	117 (102–137)	117 (102–137)	0.02	0.04	0.01	0.06	0.05
HDL-C, mg/dl	56 (49–68)	56 (50–68)	55 (49–68)	56 (51–68)	56 (50–68)	56 (49–68)	0.05	0.04	0.01	0.05	0.03
Currently smoking, n (%)	31.589 (32)	3.922 (30)	6.756 (29)	6.251 (31)	4.815 (32)	6.764 (32)	0.02	0.05	0.05	0.05	0.02
Drinking, n (%)	46.397 (47)	5.884 (45)	10.484 (45)	9.276 (46)	7.072 (47)	9.724 (46)	0.04	0.03	0.02	0.02	0.02
Diabetes events		Total population, n (%)					4.631 (4.69)
		Male, n (%)					3370 (3.41)
		Female, n (%)					1261 (1.27)
Consistency in number of reported answers for eating habits		Skipping breakfast ≥3 times/week
		Consistent, n					92.259
		Inconsistent, n					6.459
		Dinner 2 h or less before bedtime
		Consistent, n					73.292
		Inconsistent, n					25.426
		Fast eating
		Consistent, n					67.834
		Inconsistent, n					30.884
		Snacking after dinner
		Consistent, n					72.796
		Inconsistent, n					25.922
		Eating vegetables at every meal
		Consistent, n					56.665
		Inconsistent, n					42.053

BR skipping breakfast ≥ 3 times/week, DIN dinner consumed ≤ 2 h before bedtime, FAST fast eating, SNK snacking after dinner, VEG eating vegetables at every meal, SMD standardized mean difference. Data for age, BMI, currently smoking and drinking are presented as mean ± SD or n (%); data for SBP, DBP, LDL-C and HDL-C are presented as median (interquartile range). “Before matching” values reflect the full eligible cohort (N = 98,718) before habit-specific propensity score matching; “after matching” values and SMDs are reported separately for each of the five eating habits, matched independently. SMD < 0.1 was considered indicative of adequate covariate balance between comparison groups. Male and female subgroup totals may not sum to the overall matched sample size, as sex-stratified matching was performed independently using the same nearest neighbor algorithm and caliper (0.2), and not all subjects within each stratum found an eligible match within the caliper

Results

The association between each of the five eating-related lifestyle variables and the incidence of type 2 diabetes was analyzed employing Cox proportional hazards regression adjusted for all covariates described in the “Methods Section” as the entire PSM cohort, and the number for each variable population is presented in Supplemental Table 1. As shown in Fig. 1 and Table 2, among 50,180 propensity-score-matched participants, subjects who reported habitual fast eating exhibited a 12% increased risk of developing diabetes compared with those who did not eat rapidly (HR: 1.12; 95% CI: 1.02–1.23; p = 0.01). There were no statistically significant associations between diabetes incidence and skipping breakfast, having dinner close to bedtime, snacking after dinner or eating vegetables at every meal.

Fig. 1.

Hazard ratios of eating habits for new-onset type 2 diabetes in the whole population. Multivariable Cox proportional hazards model adjusted for age, sex, BMI, LDL-C, HDL-C, smoking and drinking status using PSM. PSM propensity score matching, BMI body mass index, LDL-C low-density lipoprotein cholesterol, HDL-C high-density lipoprotein cholesterol, HR hazard ratios, CI confidence interval

Table 2.

Hazard ratios of eating habits for new-onset type 2 diabetes

Variable	Overall HR (95% CI)	p	Men HR (95% CI)	p	Women HR (95% CI)	p
Skipping breakfast ≥ 3 times a week	1.11 (0.90–1.37)	0.28	1.24 (0.99–1.55)	0.05	1.21 (0.69–2.12)	0.48
Dinner ≤ 2 h before bedtime	1.10 (0.94–1.27)	0.20	1.05 (0.90–1.24)	0.49	1.42 (0.96–2.11)	0.07
Fast eating	1.12 (1.02–1.23)	0.01	1.02 (0.95–1.09)	0.54	1.14 (1.00–1.30)	0.03
Snacking after dinner	1.00 (0.82–1.21)	0.98	1.00 (0.79–1.27)	0.95	0.90 (0.65–1.26)	0.56
Eating vegetables at every meal	1.12 (0.93–1.34)	0.21	1.09 (0.89–1.34)	0.38	1.09 (0.76–1.56)	0.62

Hazard ratios (HRs) and 95% confidence intervals (CIs) were estimated using Cox proportional hazards regression models. P-values were calculated using Wald tests based on the estimated regression coefficients and their CIs

HR hazard ratios, CI confidence interval

To investigate possible differences in risk factors between the sexes, the post hoc analysis was stratified by sex. In the male population (n = 55.507), none of the eating habits were statistically significant, although skipping breakfast showed a trend toward increased risk (HR: 1.24, 95% CI: 0.99–1.55, p = 0.05; matched population in the time-restricted Cox = 9.436) (Fig. 2). In the female cohort (n = 43.211), only habitual fast eating showed a significant association with increased diabetes risk (HR: 1.14; 95% CI: 1.00–1.30; p = 0.03; matched population = 7.168), which was consistent with findings observed in the overall population (Fig. 3).

Fig. 2.

Hazard ratios of eating habits for new-onset type 2 diabetes in the male population. Multivariable Cox proportional hazard model adjusted for age, BMI, LDL-C, HDL-C, smoking and drinking status using PSM. PSM propensity score matching, BMI body mass index, LDL-C low-density lipoprotein cholesterol, HDL-C high-density lipoprotein cholesterol, HR hazard ratios, CI confidence interval

Fig. 3.

Hazard ratios of eating habits for incidence of type 2 diabetes in the female population. Multivariable Cox proportional hazards model adjusted for age, BMI, LDL-C, HDL-C, smoking and drinking status using PSM. PSM propensity score matching, BMI body mass index, LDL-C low-density lipoprotein cholesterol, HDL-C high-density lipoprotein cholesterol, HR hazard ratios, CI confidence interval

Discussion

Our present study, using a large PSM cohort, provides evidence for an association between specific eating habits and the development of type 2 diabetes. The most significant result is confirmation of fast eating as a key risk factor across the overall study population, a finding highlighted by a pronounced sex-specific effect.

Our finding that habitual fast eating was the sole statistically significant risk factor among the five habitual eating factors examined in the combined cohort (HR: 1.12; 95% CI: 1.02–1.23) is highly consistent with a growing body of literature, particularly from East Asian populations [6]. Sakurai et al. identified a similar association in a cohort of middle-aged Japanese male factory workers, in which the risk of developing diabetes over 7 years in fast eaters was nearly double that observed in slow eaters (HR: 1.97; 95% CI: 1.10–3.55) [21]. Their analysis used Cox proportional hazards models with sequential adjustment for age, family history of diabetes, smoking, alcohol consumption, habitual exercise, hypertension and dyslipidemia. In contrast, the present study used PSM based on sex, age, BMI, blood pressure, LDL and HDL cholesterol, and smoking and drinking status, before applying the Cox regression method and included both men and women. A much larger Japanese cohort further confirmed this finding. Kudo et al. drew on data from the nationwide Specific Health Check and Guidance System to examine new-onset diabetes over 3 years in 197,825 non-diabetic adults aged 40–74 years, including both men and women [6]. Using multivariable-adjusted logistic regression, they assessed four eating behaviors: fast eating, snacking after supper, snacking before sleep and skipping breakfast, with sequential adjustments for sex, age, BMI, fasting plasma glucose, smoking, drinking and weight change history. Fast eating emerged as the only habit that remained an independent predictor across all models (OR: 1.26; 95% CI: 1.20–1.33), with the association persisting after adjustment for weight gain [6]. Further corroboration comes from a large-scale retrospective cohort study of 128,594 Japanese corporate employees followed over a mean of 6.4 years, in which fast eating was similarly identified as the strongest independent behavioral predictor of incident type 2 diabetes among all eating habits examined (HR: 1.96; 95% CI: 1.73–2.21) [20]. That study employed Cox proportional hazards regression with adjustment for age, smoking, exercise habits, night shift work and alcohol intake, a covariate structure broadly comparable to the present study’s approach, although without a pre-matching step. Taken together, the convergence of findings across these methodologically distinct Japanese cohorts, including the factory-worker cohort of Sakurai et al. [21], the nationwide data of Kudo et al. [6] and the corporate cohort of Toyokuni et al. [20], provides robust, cumulative evidence that fast eating is a consistent and reproducible risk factor for type 2 diabetes development in the Japanese population. The metabolic consequences of fast eating extend beyond the well-recognized glycemic risk. Zhang et al. investigated the relationship between fast eating frequency and metabolic dysfunction-associated steatotic liver disease in 1965 Chinese adults from ten specialist clinics, using multivariable logistic regression with sequential adjustments for sex, age, smoking and drinking status, BMI, waist circumference and a comprehensive panel of clinical metabolic parameters, including HbA1c, fasting glucose, triglycerides, total cholesterol, HDL-c, LDL-c and liver enzymes [22]. Participants who ate fast at least twice a week had 81% higher odds of having metabolic dysfunction-associated steatotic liver disease than those who did so at most once a month, and this association persisted after full covariate adjustment (OR: 1.81; 95% CI: 1.26–2.59). A companion meta-analysis of five eligible studies corroborated this result, yielding a pooled OR of 1.22 (95% CI: 1.07–1.39) [22]. In contrast to the present study, Zhang et al. used a more detailed metabolic panel but relied on cross-sectional data without matching. However, the PSM design of the present study controlled for baseline confounding across groups before the survival analysis, thereby allowing stronger causal inferences regarding the temporal relationship between fast eating and metabolic disease onset.

An important finding reported by Toyokuni et al. [20] that warrants specific attention in this context is a significant interaction between fast eating and BMI category. In their cohort, the hazard associated with fast eating was substantially elevated in participants with a BMI < 25 kg/m² (HR: 1.61; 95% CI: 1.37–1.90), whereas it was no longer statistically significant in those with BMI ≥ 25 kg/m² (HR: 1.08; 95% CI: 0.89–1.30; P for interaction = 0.0007). The authors interpreted this as indicating that in individuals already carrying excess adiposity, the metabolic burden of obesity itself may outweigh the additional contribution of eating speed to glucose dysregulation. The present study’s PSM design, which balanced groups on baseline BMI, might have attenuated the detectability of such an interaction as a consideration worth acknowledging when comparing hazard ratios across studies.

The physiological mechanism underlying the impact(s) of fast eating on metabolic disorders, such as diabetes, is multifaceted and involves both behavioral and hormonal pathways. Reportedly, rapid food consumption frequently leads to a failure to recognize satiety signals, resulting in greater caloric intake and subsequent weight gain, a major driver of insulin resistance [23]. This suggests a direct metabolic consequence. Moreover, fast eating was shown to be associated with a larger and more rapid postprandial glucose excursion and a subsequent exaggerated insulin response [7]. A randomized controlled crossover trial with healthy young women as subjects showed that fast eating (10 min) versus slow eating (20 min) resulted in a significantly higher mean amplitude of glycemic excursion (3.67 vs. 2.67 mmol/l) and increased postprandial glucose peaks across all meals [24]. Over time, this chronic demand on pancreatic beta-cells could lead to their exhaustion and thereby lead to the development of overt diabetes. Furthermore, another study suggested that fast eating might trigger the release of specific pro-inflammatory cytokines, such as interleukin-1β and interleukin-6, factors known to increase systemic insulin resistance [8]. The mechanism was speculated to likely also involve impaired satiety hormone responses, with fast eating being associated with both prolonged feelings of hunger and reduced postprandial satiety [25].

An interesting finding of our current study suggests that the significant association between fast eating and incident diabetes was mainly driven by female subjects in the dataset (HR: 1.14; 95% CI: 1.00–1.30), as reflected in sex-specific differences in risk profiles. It was even suggested that, given fundamental sex-based metabolic differences, women typically exhibit better glucose tolerance than men, which might mask the effects of certain risk factors until a higher threshold of metabolic stress is reached [26]. The stronger effect of fast eating on diabetes development in women clearly warrants further investigation. Considering the multiple comparisons conducted across eating habits and sex strata, as well as the modest effect sizes observed, these findings are best regarded as hypothesis-generating rather than conclusive. This is especially relevant for the sex-stratified result, given its post hoc nature.

This sex-specific finding contrasts with the results reported by Toyokuni et al. [20], who found no statistically significant interaction between sex and eating behavior in their predominantly male cohort (76.2% male). While both studies identified fast eating as a meaningful risk factor overall, the absence of a sex interaction in Toyokuni et al. may partly reflect the substantial imbalance in their sex distribution, which could have limited statistical power to detect differential effects in women. The demographic composition of the study population, therefore, represents a meaningful methodological difference between the two investigations and may account for the divergent sex-specific conclusions. On the other hand, sex differences in gastric emptying rate, gut hormone secretion, or even the social context of eating [27] might also have contributed to the sex disparity observed in this study.

Numerous worldwide meta-analyses have consistently linked irregular meal patterns, particularly skipping breakfast, to an increased risk of type 2 diabetes [12, 13]. A comprehensive meta-analysis of 8 studies involving 106,935 subjects also found that skipping breakfast was associated with a pooled relative risk of 1.21 (95% CI: 1.12–1.32) for type 2 diabetes [12]. A Korean study showed that men were almost twice as likely to have increased insulin resistance as women when breakfast was skipped [28]. Skipping breakfast has been shown to impair insulin sensitivity at subsequent meals and has been associated with a higher risk of developing insulin resistance, especially in subjects with pre-existing metabolic risk factors [28, 29]. Additionally, disruption of circadian clock genes due to skipping breakfast may affect insulin secretion and increase HbA1c levels, particularly in obese individuals [29]. In contrast to these previous reports and some of the earlier evidence presented, the current study found that skipping breakfast was not associated with diabetes risk in a large population analysis that adjusted for potential confounders in the PSM analysis, as described above.

Whether or not having dinner shortly before bedtime is a risk factor for diabetes remains an issue of increasing importance in the field of chrononutrition, i.e., the study of how the timing of food intake interacts with the body’s circadian system [9]. The human body’s glucose tolerance reaches a nadir in the evening due to the natural circadian rhythm of insulin sensitivity [30]. Consuming a large meal late at night, close to sleep, forces the body to process a significant glucose load when its metabolic machinery is at its least efficient. This misalignment between nutrient intake and the body’s internal clock has the potential to lead to chronic hyperglycemia and increased risk of diabetes [10]. A randomized crossover study found that late dinner (22:00), compared with a more routine dinner time (18:00), resulted in higher nocturnal glucose levels, a delayed triglyceride peak and reduced fatty acid oxidation, with effects most pronounced in habitual early sleepers [31]. Moreover, the NutriNet-Santé cohort study of 103,389 adults found that later meal times were associated with higher cardiovascular disease risk, and this association was stronger in women [32], suggesting a potential sex-specific vulnerability to this chrononutritional misalignment. This observation was supported by experimental studies showing that late eating may promote a positive energy balance and increase the risk of obesity and metabolic dysfunctions by altering hormonal and gene expression [33]. Notably, consuming dinner when endogenous melatonin levels are elevated, as occurs in the late evening, may well impair glucose tolerance by reducing insulin secretion, an effect that was notably stronger in carriers of the G-allele of the melatonin receptor 1B, which is associated with type 2 diabetes [34, 35]. To date, to our knowledge, there have been neither large-scale prospective cohort studies nor retrospective analyses similar in design to our current investigation. Thus, it is meaningful that our study showed no association between dinner shortly before bedtime and diabetes risk, although a clinically relevant effect cannot be ruled out. Our data highlight the necessity of chrononutrition studies in this population.

Snacking after dinner showed no significant association with diabetes risk in the present study (HR: 1.00; 95% CI: 0.82–1.21), which contrasts with the findings of Toyokuni et al., who reported a significant association between snacking after dinner and incident type 2 diabetes (HR: 1.07; 95% CI: 1.003–1.14) in a comparably designed Japanese health checkup cohort [20]. A related Panasonic cohort study similarly linked snacking after dinner to incident cardiovascular events, a condition that shares overlapping metabolic risk pathways with type 2 diabetes [36]. This difference with our null finding may reflect differences in how snacking was operationalized; our questionnaire captured the regular occurrence of post-dinner snacking without information on snack composition, portion size or caloric content, factors likely to be more mechanistically relevant to glucose regulation than snacking frequency alone.

Contrary to some of the existing literature, daily vegetable consumption showed a non-significant trend toward increased diabetes risk in our cohort (HR 1.12, 95% CI 0.93–1.34). This finding should be interpreted cautiously given its non-significance. It is notable that our null finding is consistent with a dose-response meta-analysis of prospective cohort studies, which found that vegetable intake alone, independent of fruit intake, was not significantly associated with type 2 diabetes risk (RR: 0.98, 95% CI: 0.95–1.01 per serving/day) [37], suggesting that vegetable intake alone may be a weaker or less consistent predictor of diabetes risk than fruit and vegetable intake combined. Our habit variable also captured only the frequency of vegetable consumption without information on portion size, preparation method or overall dietary quality, all of which influence the metabolic impact of vegetable intake. Notably, an interventional study showed that the order in which vegetables are consumed within a meal, rather than the frequency of consumption alone, significantly reduces postprandial glucose and insulin excursions regardless of eating speed [38], suggesting that timing and food order, dimensions not captured by our habitual frequency-based measure, may be more metabolically relevant than whether vegetables are consumed at every meal. Reverse causation is another possibility, as participants with early or undiagnosed metabolic dysfunction may have increased vegetable intake in response to physician guidance or emerging health concerns before diagnosis. Future studies incorporating detailed dietary quality data would help clarify this unexpected finding.

Limitations

Limitations of this study must be acknowledged. First, this was not a randomized-controlled study involving a prospective cohort. Thus, PSM of a retrospectively collected dataset was used to simulate a target trial. Second, our study population was limited to Shizuoka prefecture, and the results thus might not be applicable to the entire Japanese population because of cultural variations in eating and/or living among locations. Third, eating habits were assessed using self-administered questionnaires, which captured the frequency and timing of specific behaviors without collecting detailed information on meal composition, dietary quality or specific nutrient intake such as added sugar. As macronutrient balance, energy intake, added sugar consumption and food processing all play important roles in the development of type 2 diabetes, residual confounding related to unmeasured nutritional factors cannot be ruled out. Fourth, while self-reported eating rates have shown reasonable construct validity, including a dose-response correlation with body mass index in a Japanese population [39], some misclassification is possible. Fifth, subjects receiving dyslipidemia medications were not excluded from the analysis, although statins are reportedly a risk factor [40]. Instead, lipid-related factors were accounted for through covariate adjustment and propensity score-based methods, allowing for a more representative assessment of the associations between eating habits and new-onset diabetes in a real-world population. Future studies that account for detailed dietary intake data and medication-specific information would likely further clarify these associations. Sixth, participants who changed their reported habit at any point during follow-up were excluded from analysis of that specific variable. The frequency of response changes observed (ranging from 1 to 10 times per participant) suggests this predominantly reflects instability in self-report rather than a single discrete behavior change, which prevented a meaningful comparison between stable and “changed” groups within the scope of this study. We cannot exclude the possibility that some exclusions reflect genuine health-driven behavior change, which could introduce bias in either direction. A dedicated analysis of response-change patterns and their clinical correlates represents an important direction for future research.

Conclusion

Our findings support the role of habitual fast eating as a modifiable behavioral risk factor for type 2 diabetes, with a particularly pronounced association in women. Given that eating speed is a readily observable and modifiable behavior in clinical encounters, these results provide preliminary support for considering eating speed as part of lifestyle-based diabetes prevention discussions, pending confirmation in future studies. Healthcare providers may consider raising eating pace as one of several modifiable behaviors relevant to diabetes risk, alongside established dietary guidance—not only in Japan but also in global contexts where fast eating patterns are prevalent.

Supplementary Information

Below is the link to the electronic supplementary material.

Author Contributions

Emmanuele D. S. Andrade: Conceptualization, data curation, methodology, software, formal analysis, visualization, investigation, writing – original draft and review and editing. Yuichiro Obana: Conceptualization and methodology. Mika Suzuki: Resources. Shigenori Nagai, Tetsuya Hosooka, Yuki Shimba: Supervision. Toshio Hosaka: Supervision, conceptualization, project administration, writing – review and editing.

Funding

This work was supported by the MEXT scholarship (E. D. S. A, grant no. 230464). The Rapid Service Fee was funded by the authors.

Data Availability

The datasets generated and/or analyzed during the present study are not publicly available because data sharing was not included in the consent form. However, they are available from the corresponding author upon reasonable request.

Declarations

Conflict of Interest

Emmanuele D. S. Andrade, Yuichiro Obana, Mika Suzuki, Shigenori Nagai, Tetsuya Hosooka, Yuki Shimba and Toshio Hosaka declare no conflict of interest.

Ethical Approval

This study was conducted according to the guidelines laid down in the Declaration of Helsinki, and all procedures involving research study participants were approved by the University of Shizuoka Ethics Committee (Approval no. 6–3). After approval, consent was obtained in the opt-out form by posting at the medical institution and on its website. Data were anonymized by the medical institution prior to transfer to the research team, and the data usage agreement between the medical institution and the research team specified that the data would be used for research purposes only.