figure a


People with type 2 diabetes show a variety of metabolic derangements clustered around insulin resistance, pancreatic beta cell dysfunction and systemic inflammation [1, 2]. In particular, resistance to insulin’s inhibitory action on lipolysis results in increased flux of NEFA from adipose depots to all tissues. In the liver, excess delivery and oxidation of fatty substrates stimulates ketogenesis [3], thereby limiting diet-induced fatty liver injury [4]. Ketones are transported to extrahepatic tissues, where they are taken up by monocarboxylate transporters 1 and 2 [5] and readily oxidised. Because of their favourable energetics, ketones are regarded as a thrifty fuel for organs including the heart, brain and kidneys [6,7,8].

Ketone bodies, mostly β-hydroxybutyrate (βHB), circulate at higher concentrations in individuals with diabetes as a metabolic signature of enhanced lipolysis and whole-body fat oxidation [9]. Fasting βHB levels, however, rarely exceed 1 mmol/l outside extraordinary circumstances (prolonged fasting, concomitant ketogenic diets or energy restriction, insulin omission, etc.). In non-diabetic Korean individuals, those with fasting ketonuria had a lower prevalence of obesity and the metabolic syndrome compared with those without ketonuria [10]. However, the clinical and prognostic significance of mild ketosis, as reflected by the detection of ketonuria, is still uncertain. Given the potentially beneficial role of modest increments in serum ketone bodies in energy metabolism and signalling [11], we investigated the relationship between the presence of spontaneous fasting ketonuria and the development of diabetes in a large population-based 12 year longitudinal study.


Study population

The study population comprised participants in the ongoing Korean Genome and Epidemiology Study (KoGES), a prospective, community-based survey launched in 2001 by the Korean Center for Disease Control and Prevention [12]. KoGES enrolled 10,030 participants, aged 40–69 years, and includes the results of biennial health examinations and self-reported questionnaires regarding personal and family medical history, smoking and alcohol consumption, and exercise habits. For the present analysis, we excluded current steroid users (n = 15), people with previously diagnosed diabetes and those taking any oral hypoglycaemic agents or insulin (n = 682), people who had missing results from a 75 g OGTT (n = 45) and people with a plasma glucose concentration ≥7.0 mmol/l after an overnight fast or ≥11.1 mmol/l 2 h after glucose ingestion (n = 553) (electronic supplementary material [ESM] Fig. 1). Of the remaining 8735 individuals, 31 who had missing data of urinary ketones and one with 3+ ketonuria test strip results—who might have had an abnormal physiological condition—were also excluded, leaving data on 8703 individuals available for analysis.

In addition, to explore the relationship between urinary ketones and serum βHB levels we retrospectively collected urinalysis data and serum βHB levels from individuals with type 2 diabetes treated with sodium–glucose cotransporter 2 (SGLT2) inhibitors (n = 441) at the Severance Hospital, a university-affiliated tertiary care hospital in Seoul, owing to lack of data of serum βHB in the KoGES database. All participants provided informed consent and the study protocol was approved by the Korean Center for Disease Control and Prevention and the Institutional Review Board of the Severance Hospital (IRB no. 4-2017-0287) in conformity with the Declaration of Helsinki.

Data and measurements

Participants are examined every 2 years and complete self-reported questionnaires [13]. Smoking consumption was categorised as never, past smoker or current smoker, and alcohol consumption as none, <1, 1 – <5 or ≥5 drinks per day, respectively. Exercise status was categorised as none or one or more times a week. Dietary assessment was performed at the baseline examination by well-trained interviewers using a 103 item semi-quantitative food frequency questionnaire; total energy intake and amounts and percentages of macronutrients, including carbohydrate, protein and fat, were estimated from the food frequency questionnaire [14,15,16]. The questionnaire was developed to assess the usual dietary intake of Korean adults; its validity and reproducibility were examined for the KoGES in previous reports [17, 18]. Height, body weight and waist circumference were measured and BMI was calculated; obesity was defined as a BMI ≥25 kg/m2, according to the criteria of the WHO Asia Pacific region [19]. Total body fat-free mass (kg) and body fat mass (kg) were assessed by multifrequency bioelectrical impedance analysis (InBody 3.0; Biospace, Seoul, Republic of Korea). BP was measured three times in the morning after at least 10 min in the sitting position. Hypertension was diagnosed in participants taking antihypertensive medication or with BP ≥140/90 mmHg. Blood and urine samples were obtained after at least an 8 h fast. Plasma total cholesterol, triacylglycerol and HDL-cholesterol were measured using a 747 chemistry analyser (Hitachi, Tokyo, Japan). LDL-cholesterol was calculated using the Friedewald equation [20]. Plasma glucose and insulin levels, sampled at 0 min, 60 min and 120 min during the OGTT, were measured using the hexokinase method and radioimmunoassay (LINCO kit; St Charles, MO, USA), respectively. HbA1c was determined by HPLC (Variant II; Bio-Rad Laboratories, Hercules, CA, USA). Fresh urine samples collected in the morning after the first voiding were analysed by physicians using URISCAN Pro II (YD Diagnostics, Seoul, Republic of Korea). Ketonuria results were categorised as absent, trace (50 mg/l), 1+ (150 mg/l), 2+ (400 mg/l) or 3+ (800 mg/l) based on a colour scale [10]. For the present analysis, the presence of ketonuria was defined as higher than trace level (including trace, 1+ or 2+) but did not include 3+ test strip results. For the analysis of the relationship between strip-based ketonuria and serum βHB levels, urine samples were collected in the morning and analysed using URISCAN Pro II (YD Diagnostics). Serum βHB levels were determined using an enzymatic immunoassay kit (Nittobo Medical, Tokyo, Japan).

HOMA-IR was calculated and insulin resistance was defined as HOMA-IR >2.5 [21]. The insulinogenic index (IGI) was determined by plasma insulin and glucose levels at 0 min and 60 min of OGTT [22]. Incident diabetes was defined as current use of oral glucose-lowering medication or insulin, a fasting glucose level ≥7.0 mmol/l, or a 2 h glucose level ≥11.1 mmol/l on the biennial OGTT, based on the 1997 ADA criteria [23]. The metabolic syndrome was defined according to the revised National Cholesterol Education Program definition that includes three or more of the following criteria: waist circumference >90 cm in men or >80 cm in women by the WHO Asia Pacific abdominal obesity criteria; serum triacylglycerol ≥3.9 mmol/l or medication use; HDL-cholesterol <1.0 mmol/l in men or <1.3 mmol/l in women; BP ≥130/85 mmHg or use of antihypertensive medication; or serum fasting glucose ≥5.6 mmol/l or use of glucose-lowering medication [24, 25].

Statistical analysis

To minimise potential confounding, we performed propensity score-matched analyses based on the one-to-four nearest-neighbour matching method (Greedy method) using a multivariable logistic regression model. In the propensity score model, presence of ketonuria was the dependent variable and age, sex, BMI, HbA1c, post-load 1 h and 2 h glucose, and family history of diabetes were included as covariates, yielding 185 participants with, and 740 participants without, ketonuria.

Data are presented as mean ± SD for continuous variables or frequencies for categorical variables. Group differences were tested using the unpaired Student’s t test for continuous variables and the χ2 test for categorical variables. Comparisons of ketonuria status and serum βHB levels were calculated using the Jonckheere–Terpstra trend test. Cumulative diabetes incidence was assessed by Kaplan–Meier survival functions, and the associated probability was tested using the logrank test. Cox proportional hazards analysis was performed to estimate HRs and 95% CIs, to test the independent association between ketonuria and incident diabetes. In multivariate analyses, age, sex, obesity (BMI ≥25 kg/m2 as a categorical value), post-load 1 h and 2 h glucose, post-load 1 h and 2 h insulin concentrations, insulin resistance (HOMA-IR >2.5 as a categorical value), eGFR, triacylglycerols, HDL-cholesterol, family history of diabetes, hypertension, smoking, alcohol consumption, exercise status and carbohydrate intake (g) were used as covariates. A p value <0.05 was considered statistically significant. Statistical analyses were performed using IBM SPSS Statistics version 23.0 for Windows (IBM, Armonk, NY, USA).


Clinical and laboratory characteristics of participants at baseline

The baseline characteristics of the 8703 participants of the whole cohort by presence of ketonuria are shown in ESM Table 1. The mean age of all participants (53% women) was 52 ± 9 years and their mean BMI was 24.5 ± 3.1 kg/m2. In the 185 participants with ketonuria, the distribution of ketonuria was 76.4% (trace), 21.0% (1+) and 2.6% (≥2+). Compared with participants without ketonuria, those with ketonuria were less obese, had less fat mass and were less likely to have the metabolic syndrome, but had similar fat-free mass. Levels of HbA1c, fasting and post-load 1 h insulin, HOMA-IR and triacylglycerols were significantly lower, while levels of post-load 1 h and 2 h glucose and HDL-cholesterol were higher in participants with ketonuria compared with those without ketonuria; fasting glucose concentrations were similar in both groups. Individuals with fasting ketonuria consumed slightly less total energy and carbohydrate compared with those without ketonuria, but the percentage of total energy from carbohydrate was not significantly different between the two groups (ESM Table 1). As different baseline characteristics may introduce confounding in the analysis, we used propensity score matching on age, sex, BMI, HbA1c, post-load 1 h and 2 h glucose, and family history of diabetes, in a 1:4 ratio. Table 1 shows that the baseline clinical and laboratory characteristics were comparable between the propensity score-matched groups, with the exception of fasting and post-load 2 h insulin, triacylglycerol concentrations and total energy and carbohydrate intake, which were lower, and HDL- and LDL-cholesterol, which were higher, in the ketonuria compared with the no-ketonuria group.

Table 1 Baseline characteristics of the study participants according to the presence or absence of ketonuria after propensity score matching (n = 925)

Association between ketonuria and incident diabetes

Of the 925 participants in the propensity score-matched cohort, 190 (20.5%) developed diabetes during the 12 year follow-up period, a crude incidence rate of 1.7% per year. Of these 925 participants, 165 (22.3%) without ketonuria and 25 (13.5%) with ketonuria developed diabetes. In the whole cohort, 1450 (17.0%) participants without ketonuria and 30 (15.4%) with ketonuria developed diabetes. As shown in Fig. 1, participants with ketonuria had a significantly lower cumulative incidence of diabetes compared with those without ketonuria (p = 0.024 by logrank test); by Cox proportional hazards analysis, the relative risk reduction was 0.64 (95% CI 0.42, 0.98; p = 0.040) after adjustment for age, sex and obesity (model 1, Table 2). Also adjusting for post-load 2 h glucose concentrations, post-load 1 h and 2 h insulin concentrations, HOMA-IR, triacylglycerols, HDL-cholesterol and eGFR (model 2, Table 2) did not change the association of ketonuria with incident diabetes, which was maintained at the same strength after further adjustment for clinical variables including family history of diabetes, hypertension, smoking, alcohol consumption and exercise status (model 3, Table 2; Fig. 2). After additional adjustment for carbohydrate intake, the association between fasting ketonuria and incident diabetes remained statistically significant (HR 0.63; 95% CI 0.41, 0.97; p = 0.037) (model 4, Table 2).

Fig. 1
figure 1

Twelve year diabetes-free survival according to baseline ketonuria; p = 0.024 by logrank test

Table 2 Association of urinary ketones with the incidence of diabetes in Cox models in the propensity score-matched sample (n = 925)
Fig. 2
figure 2

Multivariate-adjusted Cox regression analysis for incident diabetes by ketonuria status in the propensity score-matched sample (n = 925)

Analysis of the whole cohort (n = 8703) yielded very similar results, i.e. an inverse association between ketonuria and incident diabetes, with an HR of 0.66 (95% CI 0.45, 0.96) after adjusting for the same set of variables as used in the propensity score-matched groups (ESM Table 2, ESM Table 3). Of note, in this full model most classical factors associated with diabetes (except for sex, drinking habits and eGFR) were statistically significant predictors of its incidence (ESM Table 3).

Trends in metabolic variables during 12 years by ketonuria status

The longitudinal changes in metabolic variables (body composition, glucose, HOMA-IR and IGI) were analysed by presence of ketonuria in the whole cohort and in the propensity score-matched subcohort (Fig. 3; ESM Fig. 2). In the whole cohort, fasting and post-load 2 h glucose levels and HOMA-IR went upwards to a similar extent in the two groups, while BMI, fat mass and per cent body fat remained lower in the participants with ketonuria compared with those without ketonuria (ESM Fig. 2). In the propensity score-matched cohort, no significant differences were observed in body weight and composition between groups over time (Fig. 3). Fasting glucose levels increased similarly between the two groups, whereas post-load 1 h and 2 h glucose, which were superimposable at baseline, were maintained at lower levels in the ketonuria group compared with the no-ketonuria group. Whereas HOMA-IR showed a gradual increase without significant difference between the groups, participants with ketonuria had a higher IGI.

Fig. 3
figure 3

Longitudinal changes of glycometabolic variables according to presence of ketonuria in the propensity score-matched cohort: (a) fasting glucose, (b) post-load 1 h glucose, (c) post-load 2 h glucose, (d) HOMA-IR, (e) IGI, (f) BMI, (g) fat mass, (h) body fat per cent, (i) fat-free mass. Plots are mean ± SEM. *p< 0.05, **p< 0.01 between two groups

Relationship between urinary ketones and serum βHB level

The distribution of urinary ketones in 441 individuals with type 2 diabetes treated with SGLT2 inhibitors is shown in ESM Fig. 3. The majority (n = 275, 62.4%) were negative for strip-based ketonuria. Individuals with trace, 1+ and ≥2+ ketonuria comprised 23.8%, 9.3% and 4.5%, respectively. As shown in ESM Fig. 4, serum βHB levels gradually increased in individuals with trace, 1+ and ≥2+ ketonuria (mean serum βHB levels: none, 0.13 ± 0.09 mmol/l; trace, 0.29 ± 0.29 mmol/l; 1+, 0.57 ± 0.49 mmol/l; ≥2+, 2.27 ± 1.70 mmol/l; p for trend <0.001).


The main finding of the study was that ketonuria, as detected by clinical-grade strips in the morning after an overnight fast, was present in a small percentage (2.2%) of a non-diabetic population-based cohort, and was associated with a reduced rate of incident diabetes during 12 years of follow-up. The relative diabetes risk reduction was consistent in both size and statistical significance across the whole cohort and a propensity score-matched subcohort, and resisted multivariate adjustment in both datasets. In a full multivariate model including the expected diabetes predictors (age, obesity, familial diabetes, hypertension, insulin resistance as HOMA-IR, serum lipids and plasma glucose levels), the presence of ketonuria was still an independent inverse predictor of incident diabetes (ESM Table 3). In the whole cohort, ketonuric individuals were leaner, with less fat mass but similar fat-free mass, had a better serum lipid profile and lower insulin levels but higher post-OGTT glucose levels (ESM Table 1). A finer profile of the ketonuric phenotype emerges from the propensity score-matched subcohort, consisting of just lower insulin levels (both fasting and post-OGTT) and more favourable serum lipid levels (lower triacylglycerols and higher HDL-cholesterol, with marginally higher LDL-cholesterol), or, in the aggregate, a lower prevalence of the metabolic syndrome. Importantly, at the biennial visits the propensity score-matched cohort showed consistently lower post-load 1 h and 2 h glycaemias, coupled with some improvement in the IGI (Fig. 3). This finding supports the interpretation that the detection of sporadic ketonuria does identify a stable phenotype, namely, one of better insulin sensitivity (as reflected by HOMA-IR) and fewer features of the metabolic syndrome, possibly related to a healthier lifestyle (e.g. 32% vs 26% of people exercising at least once weekly; Table 1).

Several human studies have shown the influence of a low-carbohydrate, ketogenic diet on body weight and metabolism. Starvation and a ketogenic diet upregulate ketogenesis enzymes in the liver, in particular 3-hydroxy-3-methylglutaryl coenzyme A synthase 2 (HMGCS2), leading to increased ketone concentrations in urine and blood [7, 26]. Ketogenic diets also induce greater decreases in triacylglycerols and increases in HDL-cholesterol levels, and greater weight loss, compared with a low-fat diet [27]. Other ketogenic diet trials reported improved insulin resistance and BP [28], as well as improved steatosis, inflammation and fibrosis in non-alcoholic fatty liver disease [29]. Moreover, low energy intake or longer duration of fasting may play a role in the association between spontaneous ketonuria and lower risk of diabetes [30]. In our study, lower energy and carbohydrate intakes were observed in individuals with spontaneous ketonuria compared with those without ketonuria. Previously, prolonged or intermittent fasting was associated with reductions in body weight, fasting glucose and insulin levels and insulin resistance [31]. In addition, longer nocturnal fasting intervals could ameliorate systemic inflammation and positively influence other lifestyle behaviours such as sleep, which could subsequently reduce risks of chronic diseases [32, 33]. Prolonged fasting periods could alter gut permeability and composition of gut microbiota, which may lead to improvement in systemic inflammation and metabolic variables [34]. Therefore, the role of intermittent fasting and low-carbohydrate intake on the presence of spontaneous ketonuria needs to be investigated in future controlled studies.

Of relevance is that most ketogenic, low-carbohydrate diets implement a reduced energy intake and, consequently, result in some weight loss. For example, a recent non-randomised trial of a type 2 diabetes care model involving adherence to a supervised very-low-carbohydrate intake (<30 g/day) led to a marked improvement in glycaemic control at 1 year [29]. Individually monitored blood ketone levels doubled over this period, but body weight decreased by an average of 14 kg vs 1 kg in a usual care group. Thus, the relative contribution of hyperketonaemia and weight loss to the observed changes in metabolic status cannot be dissected out with this approach. By contrast, in our ketonuric participants, body weight and composition were remarkably stable over a long period of time, thereby suggesting a weight-independent link between relative hyperketonaemia and evolution of glucose tolerance.

There are several potential mechanisms for the pleiotropic role of ketone bodies in glucose metabolism. First, in the process of formation of ketone bodies, nuclear receptor peroxisome proliferator-activated receptor, alpha (PPARα) and one of its downstream targets, fibroblast growth factor 21 (FGF21), are strongly induced, and Hmgcs2 transcription is increased in the liver [35,36,37,38]. HMGCS2 may also induce Fgf21 gene expression [39]. In turn, PPARα and FGF21 play critical roles in glucose metabolism and may have preventive potential for diabetes [40,41,42]. Conversely, impaired ketogenesis in human non-alcoholic fatty liver disease/non-alcoholic steatohepatitis directs acetyl-CoA to activate tricarboxylic acid flux, thereby increasing reactive oxygen species-mediated injury, de novo synthesis of cytotoxic lipid species, and limited NADH re-oxidation [43]. Ketone bodies possess signalling activities and may be directly involved in insulin action. In hepatocytes, βHB activates AMP kinase and peroxisome proliferator-activated receptor, gamma, coactivator 1, alpha (PGC-1α), which could relate to increased liver insulin sensitivity [44, 45]. As an endogenous histone deacetylase inhibitor, βHB regulates the expression of gluconeogenic genes [7, 46, 47]. In rodents, increased blood βHB stimulates brown adipose tissue, lowers visceral and omental adipocyte size and improves insulin resistance [48, 49]. In athletes during exercise, ketosis induced by the administration of a ketone ester enhanced metabolic flexibility, reducing glycolysis and increasing intramuscular fat oxidation [50]. Through G protein-coupled receptors, βHB was reported to suppress sympathetic nervous system activity [51]. In addition, liver-derived βHB [52] or astrocyte-derived ketogenesis within the ventromedial hypothalamus may play a role in appetite regulation [53]. With regard to effects on pancreatic beta cells, it has long been known that ketone bodies can stimulate insulin secretion in experimental animals and humans [54,55,56]. Finally, the activity of βHB may reflect its potential in the regulation of inflammation and oxidative stress [57,58,59,60]. Clearly, the target organs, circumstances, time course and quantitative relevance of these mechanisms to metabolic homeostasis remain to be assessed in clinical studies.

The present study has several limitations. First, ketonuria serves as a semi-quantitative indicator of ketosis by detecting acetoacetate in the urine [61]. Second, blood βHB levels were not available in the study cohort, although the ancillary study in ESM Fig. 4 does provide a robust relationship between urine and blood ketones. Third, we assessed the dietary information on total energy and carbohydrate intake but could not assess specific types in carbohydrates, glycaemic index or glycaemic load, which could influence circulating blood glucose levels and glucose metabolism. In addition, blood samples were collected after at least an 8 h fast, but individual fasting duration, which could impact concentrations of blood and urine ketone bodies, was not available. Future research regarding the role of intermittent fasting and specific carbohydrate intake is necessary to determine the physiological mechanisms by which ketone bodies and metabolic health are associated. Moreover, in this study only baseline data of urinary ketones could be analysed, as longitudinal data had too many missing values. Finally, this study was based on a sample of Korean individuals in a population-based setting, which may limit the ability to generalise our results to other ethnicities or settings.

In conclusion, we show that spontaneous fasting ketonuria is significantly associated with a reduced risk of developing type 2 diabetes, after adjustment for multiple covariates. Spontaneous fasting ketonuria may be a novel signature in the modulation of glucose metabolism and have the potential to prevent diabetes.