Point-of-Care Capillary Blood Ketone Measurements and the Prediction of Future Ketoacidosis Risk in Type 1 Diabetes Free

Estimated to result in five to seven hospitalization events per 100 person-years, diabetic ketoacidosis (DKA) is a common but preventable life-threatening complication of type 1 diabetes (T1D) (1,2), and identification of those at risk is challenging. Rather than during illness while DKA is developing, we aimed to determine if levels of routine point-of-care capillary blood ketones could predict DKA months before the event. We conducted an analysis of data from the placebo group of a large-scale clinical trial of an adjunct-to-insulin medication hypothesized to increase DKA risk, because the protocol included regular weekly surveillance of fasted capillary ketones for all participants.

We analyzed previously collected prospective data from the Empagliflozin as Adjunctive to Insulin Therapy (EASE) program, which included two phase 3 randomized controlled trials of the once-daily oral sodium–glucose linked transporter (SGLT) inhibitor empagliflozin in T1D (3). Data were accessed through public data request (Vivli access platform). The study population was all 484 placebo-assigned participants. The EASE-2 trial (clinical trial reg. no. NCT02414958, ClinicalTrials.gov) had a 12-month follow-up duration, whereas the EASE-3 trial (clinical trial reg. no. NCT02580591) had 6 months. The two studies shared inclusion and exclusion criteria and protocols, and they were reported together (3). Before initiation of the study drug, patients underwent a 6-week investigator-guided insulin therapy optimization period followed by a 2-week placebo run-in period, providing an 8-week baseline period.

The diagnostic index test was capillary ketone levels (betahydroxybutyrate [BHB]) measured using an ambulatory point-of-care device (Abbott Freestyle Precision/Optium Neo; BHB detection range 0.0–8.0 mmol/L; 95% CI for standard deviation 0.02–0.05 mmol/L) (4) via electronic logbook in the baseline period before the outcome could occur. Participants were advised to test fasting capillary ketone levels two to three times per week regardless of symptoms. For analysis, ketone levels were defined by calculating the maximum, mean, or last ketone value over the 8-week baseline period.

The outcome was DKA adjudicated by a blinded clinical event committee, classified as certain or potential DKA (Supplementary Fig. 2). Certain DKA required the presence of acidosis and ketosis, along with clinical manifestations in keeping with DKA, whereas potential DKA required the presence of acidosis or ketosis, along with clinical manifestations. These outcomes were combined for analysis and are detailed in Supplementary Fig. 2 and Supplementary Table 1 (3).

Areas under curve (AUCs) for the receiver operating characteristic curves were generated for the three ketone parameters (maximum, mean, or last) in predicting future DKA risk. Diagnostic thresholds were determined using different strategies: 1) optimal thresholds simultaneously maximizing sensitivity and specificity, 2) a threshold qualitatively favoring sensitivity, 3) a threshold qualitatively favoring specificity, and 4) an existing clinical threshold for capillary ketone level elevation in individuals with T1D compared with those without T1D (values ≥0.6 mmol/L). Composite prediction rules were also explored (5). Data were processed and analyzed using R software (version 4.0.2). To determine if using all three ketone parameters could improve performance beyond using a single parameter, we used supervised machine learning methods (gradient-boosted trees [GBTs]).

Twelve participants had the outcome (certain DKA n = 6; potential DKA n = 6), and 472 did not. Participants were 53% female and age 43 ± 3 years; they had a mean BMI of 28 ± 5 kg/m² and mean HbA_1c of 8.2 ± 0.6%, with no differences between those with and without incident DKA during follow-up (Table 1). The distributions of maximum and mean capillary ketone values over the course of the baseline periods were significantly higher in those with future incident DKA compared with in those without; however, last capillary ketone values did not differ (Table 1).

Receiver operating characteristic curves for capillary ketone value parameters to predict future DKA are shown in Fig. 1A. The AUCs for the maximum, mean, and last capillary ketone value parameters during the baseline period were 0.77 (95% CI 0.66–0.88), 0.75 (0.65–0.86), and 0.64 (0.46–0.83), respectively, with no significant differences between AUCs (P = 0.29 for comparison between maximum and last parameters). For the maximum capillary ketone value, selection of a threshold value that favored specificity (≥0.8 mmol/L) had a sensitivity of 64%, a specificity of 78%, and better likelihood ratios (LRs; positive LR 2.9; negative LR 0.5) compared with other thresholds (Table 2). For the mean and last ketone values, the predictive diagnostic performance was generally similar in magnitude to the maximum ketone level in the baseline period (Table 2).

To determine if the use of maximum, mean, and last ketone levels simultaneously could improve prediction, the GBT machine learning model resulted in an AUC of 0.88 (95% CI 0.82–0.94) (Fig. 1B). Use of an optimal threshold that optimized both sensitivity and specificity resulted in a sensitivity of 91%, a specificity of 77%, and positive and negative LRs of 4.0 and 0.1, respectively (Table 2). Feature importance analysis indicated that both maximum and mean ketone values contributed significantly to the overall GBT model, whereas the last ketone value contributed less. The GBT model outperformed the multivariable logistic regression model (AUC 0.80; P = 0.046 for comparison) (Supplementary Fig. 1 and Supplementary Table 1), which was an alternate approach to using single parameter summaries of the ketone values.

In current practice, clinical measurement of capillary blood (or urinary) ketone levels is recommended when two conditions are met: 1) presence of overt illness and 2) simultaneous hyperglycemia (capillary glucose level generally exceeding 14 mmol/L or 250 mg/dL) (6–9). From the normal distribution of ketone levels in those without diabetes, the presence of a clinical threshold for capillary blood ketone (BHB) abnormality of ≥0.6 mmol/L with simultaneous hyperglycemia calls for increased oral fluids, additional insulin, repeat ketone testing, and acute care for additional testing if capillary levels, for example, exceed 1.5 mmol/L (8,10). However, the distribution of ketone levels is poorly understood when blood glucose levels are not elevated and when symptoms of overt illness are not present. This analysis provides the novel proof of concept that variation in the levels within the normal distribution of ketones in those who are not overtly ill with hyperglycemia are informative for the risk of developing DKA in the future. For example, during regular fasted periods without illness, a maximum ketone level ≥0.8 mmol/L was associated with nearly a threefold likelihood of 6- to 12-month DKA risk. This threshold value is well below the level of >1.5 mmol/L that during illness with simultaneous hyperglycemia warrants sick-day management protocols, including additional testing in acute care (8). Findings imply that such testing in all individuals with T1D, or at least in those with risk factors, could improve DKA prevention. In addition to simple strategies, such as monthly testing of fasted capillary ketone levels, automated algorithms could be implemented in ketone meters, linked mobile apps, or, in the future, continuous ketone monitoring devices to further improve prediction (11,12).

These findings advance the understanding and prediction of DKA risk. Currently, general risk factors are well recognized, including female sex, younger age, use of insulin pump therapy, lower socioeconomic status, carbohydrate restriction, and indicators of challenges with self-management skills, including history of DKA, higher HbA_1c levels, insulin dose omission, and alcohol and drug use (13–15). Despite the clear association with DKA risk, these risk factors do not provide diagnostic utility for the identification of specific individuals at highest risk.

We acknowledge limitations to our study. This represents a proof of concept within a limited sample size of DKA events. We acknowledge a possible selection bias in that trial participants are more likely to adhere to the routine capillary ketone testing protocols and routine measurements performed in the absence of illness or substantial hyperglycemia introduce additional cost, complexity of care, and self-management burden. Sufficient data for a formal cost-effectiveness analysis are currently not available. However, we wish to highlight that current guidelines call for individuals living with T1D to maintain nonexpired ketone testing strips in the event of intercurrent illness (16), and routine measurements could therefore provide a purpose for using strips that frequently would otherwise expire. Although LRs for prediction are as high as 4 for the machine learning model, we recognize that the absolute risk of DKA and predictive values for DKA are low. With a unique trial population, we were unable to validate these findings in an independent data set, but other trial programs of SGLT inhibition have been conducted that have included placebo groups that may be suitable for validation of the concept (17–20). We recognize the risk of overfitting in our machine learning models. Further work is needed to identify which specific interventions, such as optimization of basal insulin doses and supplemental diabetes self-management education, are effective in preventing DKA events.

In conclusion, routine point-of-care capillary ketone monitoring provides an additional tool for clinicians to identify those with T1D at the highest risk of future DKA events and may provide an opportunity to select individuals for clinical interventions for DKA prevention. This risk prediction feature of ketone monitoring represents a second purpose above and beyond the traditional use of ketone monitoring for sick-day management to specifically determine if DKA is developing acutely. Future research should evaluate the feasibility of the implementation of ketone monitoring using capillary as well as novel continuous combined interstitial glucose and ketone monitoring technologies (12) in unique populations, such as those with T1D using adjunctive-to-insulin SGLT inhibition, in which DKA risk is higher.

Acknowledgments. D.M. and B.A.P. are patient partners with lived T1D experience. The authors appreciate that the funder Boehringer Ingelheim made the clinical trial data available and are grateful to the entire EASE program research team, comprising 132 international research sites, and the study participants.

Funding. This publication is based on research using data from data contributor Boehringer Ingelheim, funder of the EASE program, that have been made available through Vivli, Inc. This study was supported by Diabetes Canada (operating grant OG-3-21-5572-BP), the Data Sciences Institute at the University of Toronto, and the Menkes Fund.

Vivli, Inc., has not contributed to or approved, and is not in any way responsible for, the contents of this publication.

Duality of Interest. M.F. is a consultant for ProofDx, a startup company that is developing a point-of-care diagnostic test for COVID-19 using CRISPR. D.Z.I.C. has received consulting fees or speaking honoraria or both from Janssen, Boehringer Ingelheim, Eli Lilly, AstraZeneca, Merck, and Sanofi and operating funds from Janssen, Boehringer Ingelheim, Eli Lilly, AstraZeneca, and Merck. L.E.L. received support from a Canadian Institutes of Health Research Canada graduate scholarship doctoral award. B.A.P. has received speaker honoraria from Abbott, Medtronic, Insulet, and Novo Nordisk and research support to his research institute from Boehringer Ingleheim and Bank of Montreal and has served as a consultant to Boehringer Ingelheim, Abbott, Novo Nordisk, and Vertex. No other potential conflicts of interest relevant to this article were reported.

Author Contributions. C.S., S.D., and L.E.L. completed the statistical analysis. C.S. and B.A.P. wrote the first draft of the manuscript. L.E.L. and B.A.P. developed the research question. All authors contributed to data interpretation and manuscript preparation. B.A.P. is the guarantor of this work and, as such, had full access to all the data in the study and takes responsibility for the integrity of the data and the accuracy of the data analysis.

Prior Presentation. Limited aspects of this analysis were presented at the 2022 Diabetes Canada/Canadian Society of Endocrinology and Metabolism Professional Conference, Calgary, Alberta, Canada, 9–12 November 2022, and the International Diabetes Federation Congress 2022, Lisbon, Portugal, 5–8 December 2022.