Development and validation of a multivariable risk prediction model for identifying ketosis‐prone type 2 diabetes

Abstract Background To develop and validate a multivariable risk prediction model for ketosis‐prone type 2 diabetes mellitus (T2DM) based on clinical characteristics. Methods A total of 964 participants newly diagnosed with T2DM were enrolled in the modeling and validation cohort. Baseline clinical data were collected and analyzed. Multivariable logistic regression analysis was performed to select independent risk factors, develop the prediction model, and construct the nomogram. The model's reliability and validity were checked using the receiver operating characteristic curve and the calibration curve. Results A high morbidity of ketosis‐prone T2DM was observed (20.2%), who presented as lower age and fasting C‐peptide, and higher free fatty acids, glycated hemoglobin A1c and urinary protein. Based on these five independent influence factors, we developed a risk prediction model for ketosis‐prone T2DM and constructed the nomogram. Areas under the curve of the modeling and validation cohorts were 0.806 (95% confidence interval [CI]: 0.760–0.851) and 0.856 (95% CI: 0.803–0.908). The calibration curves that were both internally and externally checked indicated that the projected results were reasonably close to the actual values. Conclusions Our study provided an effective clinical risk prediction model for ketosis‐prone T2DM, which could help for precise classification and management.

• Ketosis-prone T2DM is a specific but common subtype in newly diagnosed T2DM patients in China, with specific clinical characteristics.
• Based on age, fasting C-peptide, free fatty acids, glycated hemoglobin A 1c and urinary protein, our study provided an effective clinical risk prediction model for ketosis-prone T2DM, which could help for precise classification and management.

| INTRODUCTION
In 2019, the World Health Organization published an update on the classification of diabetes, which proposed ketosis-prone type 2 diabetes mellitus (T2DM), distinguished from traditional type 1 diabetes mellitus and T2DM. 1 Ketosis-prone T2DM is a novel form of diabetes that begins with ketosis or ketoacidosis, without evident triggers.At the time of presentation, there is a temporary secretory defect of β-cells, but it later goes into remission and does not require insulin treatment, which is significantly different from the treatment and prognosis of traditional T2DM. 2,3Ketosis-prone T2DM is now widely recognized in several ethnic populations, such as African-Americans, African-Caribbeans, and sub-Saharan Africans. 4,57][8][9] In China, the incidence of T2DM with ketosis onset is increasing annually. 10Some studies have summarized the clinical characteristics of ketosis-prone T2DM and tried to find indicators for diabetes classification. 7,11owever, the data shown in these studies were inconsistent, the indicators did not involve complete diabetes complications and other endocrine comorbidities, and did not provide auxiliary tools for differential diagnosis.
Therefore, we aim to explore the clinical characteristics of ketosis-prone T2DM from detailed inpatient medical records, analyze the related independent risk factors, and build a comprehensively prediction model for precise classification and management.

| Ethics statements
This research was carried out in conformity with the principles of the Helsinki Declaration, which was approved by the Ethics Committee of Zhejiang Provincial People's Hospital (Ethics Approval Number: 2021QT398).Informed consent was exempted, due to the retrospective nature of the data acquisition.

| Study design and population
The purpose of this cross-sectional research was to study the clinical characteristics and develop a prediction model for ketosis-prone T2DM.A total of 964 participants with T2DM satisfied the requirements for inclusion in the study.All participants included were from Zhejiang Provincial People's Hospital.The inclusion criteria were as follows: (a) newly diagnosed T2DM with/without ketosis (urine ketone ≥30 mg/dL with hyperglycemia) or ketoacidosis (arterial blood pH <7.30 or serum bicarbonate <18 mmol/L apart from ketosis and hyperglycemia) 1 ; (b) negative for glutamate decarboxylase autoantibodies, islet cell antibodies, insulin autoantibodies, and protein tyrosine phosphatase autoantibodies; and (c) no history of treatment with hypoglycemic drugs or insulin.The exclusion criteria were as follows: (a) obvious precipitating causes for the development of ketosis, such as infection, stress, surgery, trauma, excess intake of soft drinks, and infusion of a large amount of glucose; and (b) type 1 diabetes, latent autoimmune diabetes in adults, maturity onset diabetes in the young, mitochondrial diabetes, monogenic diabetes, gestational and secondary diabetes.Participants in the modeling cohort were recruited from January 2015 to December 2016, and those in the validation cohort were recruited from January 2020 to December 2021.The workflow of participants screening is shown in Figure 1.
All participants underwent physical, laboratory, and imaging examinations and were treated with intensive insulin therapy.The participants with ketosis or ketoacidosis were included in the ketosis-prone T2DM group.

| Data collection
We collected detailed information on the participants' general characteristics, such as gender, age, smoking, drinking, weight loss in the last 6 months, the onset pattern of diabetes, family history of diabetes, diabetes duration, everyday risk factors, past health records, and medication histories.Participant's height, weight, systolic blood pressure, and diastolic blood pressure were measured by the same nurse on the first day of admission.To determine body mass index, divide weight by square of height.Lower-limb atherosclerosis and carotid atherosclerosis were evaluated using a high-end color Doppler ultrasound scanner (EPIQ7, Philips).Cerebral infarction was analyzed using 3.0 T magnetic resonance imaging scanner (Discovery MR 750, GE).Peripheral neuropathy was evaluated using six-channel electromyography/evoked potential (keypoint).

| Definitions
A regimen of intensive insulin therapy comprised the use of an external insulin infusion pump or at least three daily insulin injections.Time for FPG to reach target was defined from the date of admission to the day of returning to normoglycemia or discharge.The proportion of 10-year high atherosclerotic cardiovascular disease (ASCVD) risk was evaluated using ASCVD risk assessments from the 2020 Chinese Guideline on the Primary Prevention of Cardiovascular Diseases. 12

| Statistical analysis
The Kolmogorov-Smirnov test was used to determine the index's distribution.Means and SDs were used to represent normally distributed data, and nonnormally distributed data were descriptive by medians and quartiles.Counts and proportions were used to describe categorical variables.To distinguish any differences between the two groups, the Student t test or the nonparametric Mann-Whitney U test were used.The discrepancies in categorical variables were compared through the χ 2 test.Stepwise regression analysis was used to calculate p values for trend ( p-trend) to determine the independent association between the incidence of ketosis-prone T2DM and different HbA 1c levels.Model 1 was gender and age adjusted.Model 2 was additionally modified for FFAs and FCP.Univariate logistic regression analyses were used to identify independent risk factors, and multivariate analysis was carried out using logistic regression analysis with variables that were meaningful in univariate testing included (p < .05).Forest plots of hazard ratios were plotted based on the aforementioned data.Using the nomogram function from the rms library, we created a nomogram to predict ketosis-prone T2DM.The score's accuracy was determined by calculating the receiver operating characteristic curve and analyzing the area under the receiver operating characteristic curve (AUC).The modeling cohort was compared to the validation cohort by assessing the concordance index and calibration curves on the test set to verify the performance of the model.SPSS, version 25.0 (IBM Corp.) and R, version 3.6 were used for all analyses.Differences were considered statistically significant at p < .05.

| Baseline characteristics of the study population
In total, 964 participants were included in the analysis, with 610 participants in the modeling cohort and 354 participants in the validation cohort.The incidence rate of ketosis-prone T2DM was 21.0% (128 participants) and 18.9% (67 participants), respectively.According to the χ 2 analysis, the incidence rate was not significantly different (p > .05).Baseline characteristics of participants enrolled in both studies were broadly similar.(Table 1).
The clinical characteristics of ketosis-prone T2DM are shown in Table 2. Ketoacidosis accounted for 23.4% of ketosis-prone T2DM.Obviously, participants with ketosis-prone T2DM were more likely to discovered because of the symptom such as thirst, polyuria etc. at the onset pattern of diabetes and presented with more weight loss (all, p < .001).These participants were markedly younger (42.50, interquartile range [IQR], 33.00-52.00),had higher levels of FFAs, HbA 1c , urinary protein, longer duration of intensive insulin therapy, and time for FPG to reach target, whereas those levels of HDL-C, FCP, 2-h-PCP, ALT, and BUN were significantly reduced (all, p < .05).They were more likely to require insulin therapy at discharge (p < .001).Other clinical data were not significantly different between the two groups (all, p > .05).

| Associations of different levels of HbA 1c with ketosis-prone T2DM
The associations between different HbA 1c levels and ketosis-prone T2DM with multistep adjustments are shown in Table 3.The likelihood of ketosis-prone T2DM tended to be elevated with an increase in the HbA 1c level (p-trend<.001).The results presented herein indicated that the increasing trend persisted even after correction for gender and age (p-trend<.001).When further adjusting for FFAs and FCP, the growing trend remained independently associated with HbA 1c levels ( p-trend = .001).

| Ketosis-prone T2DM with diabetes complications and endocrine comorbidities
A comparison of diabetes complications is shown in Table 4. Evidently, the incidence of diabetes complications in the ketosis-prone T2DM group was much lower, regardless of vascular disease or nephropathy.There were obvious differences in lower limb atherosclerosis, carotid atherosclerosis, and the proportion of 10-year high ASCVD risk (all, p < .05).At the same time, the differences in thyroid and bone metabolism are also described (Table 4).Levels of β-CTX was significantly higher in the ketosis-prone T2DM group (p = .010).In contrast, FT3 levels were considerably higher in the nonketotic T2DM group ( p = .001).Furthermore, other variable distributions did not correlate between the two groups (all, p > .05).Note: Results are presented as median (quartile 1 À quartile 3) or positive/negative number (positive percent).Abbreviations: 25-OHD, 25-hydroxy vitamin D; ACR, microalbumin-to-creatinine ratio; ASCVD, atherosclerotic cardiovascular disease; β-CTX, β-C terminal cross-linking telopeptide of type I collagen; FT3, free triiodothyronine; FT4, free thyroxine; P1NP, N-terminal propeptide of type I procollagen; PTH, parathyroid hormone; T2DM: type 2 diabetes mellitus; TSH, thyroid-stimulating hormone.

| Findings of univariate and multivariate logistic regression analysis
0.533-0.896,p = .005)were independent protective factors for the incidence of ketosis-prone T2DM (Table 5, Figure 2).A model for predicting the incidence of ketosis-prone T2DM was constructed based on the aforementioned results.

| Construction and evaluation of the prediction model
We created a nomogram to predict ketosis-prone T2DM (Figure 3).The probability of ketosis-prone T2DM was   This study revealed that ketosis-prone T2DM was common (20.2%) in newly diagnosed T2DM, characterized by younger age, poorer islet function, severe hyperglycemia and dyslipidemia, higher levels of FFAs, HbA 1c , urinary protein, longer duration of intensive insulin therapy and time for FPG to reach target, and lower proportion of macrovascular disease.Based on independent influence factors such as age, FCP, FFAs, HbA 1c , and urinary protein, we developed the first predictive model for ketosisprone T2DM, with an excellent discriminatory power of 0.806 while maintaining adequate calibration.
Since the 1980s, it has been recognized that some participants with ketoacidosis may have a clinical course more like that of T2DM than type 1 diabetes mellitus, which was called ketosis-prone T2DM. 2,13Recently, with the increasing awareness of ketosis-prone T2DM, the morbidity has gradually increased. 14Zhang et al reported that ketosis-prone T2DM participants accounted for 7.6% of the participants with diabetes requiring admission to a large urban hospital in China. 15Lontchi-Yimagou et al found 28.3% ketosis-prone T2DM of the hyperglycemic crises. 16We observed 20.2% ketosis-prone T2DM in newly diagnosed T2DM in our center.The discrepancy might be caused by the various types of participants recruited.As a Grade-A tertiary referral hospital in local region, our center will admit more participants with severe conditions.
In our study, we explored the clinical characteristics of participants at the onset of ketosis-prone T2DM.First, ketosis-prone T2DM participants were diagnosed at lower age and predominantly middle-aged men, which was consistent with previous studies. 2,8Genetic factors, changes in hormone levels, and environmental factors might all interact to cause the condition.Second, ketosisprone T2DM had a temporary secretory defect of β-cells.Mauvais-Jarvis et al 17 suggested that the ketosis-prone T2DM included one major subtype with preserved insulin secretion and a rarer subgroup with permanent insulin dependence.Our participants showed lower levels of FCP at the onset.Third, glucose metabolic disturbance was more severe in ketosis-prone T2DM participants, which was in line with previous studies. 3,8Our study provided further evidence of the relationship between HbA 1c levels and the development of ketosis-prone T2DM.Fourth, dyslipidemia was evident in ketosis-prone T2DM participants.The role of lipotoxicity in the β-cell failure in ketosis-prone T2DM participants was debatable.Some studies have suggested that lipotoxicity inhibited insulin gene expression 18 whereas Umpierrez et al discovered that intralipid infusion for 48 h was not affiliated with β-cell decompensation in ketosis-prone T2DM participants. 19Other teams used metabolomic approaches to demonstrate that ketosis was due to decreased ketolysis rather than overproduction of lipid. 20,21In our study, a markedly higher level of FFAs and lower HDL-C levels were seen.There was no significant difference in other lipid parameters.][25] In our study, we further explored diabetes complications and endocrine comorbidities in ketosis-prone T2DM participants.First, ketosis-prone T2DM participants had a lower rate of combined macrovascular disease at the time of onset.This might be that participants with ketosis-prone T2DM see their doctor earlier than those without due to ketosis symptoms.However, Li's team considered that the prevalence of carotid atherosclerosis or lower limb atherosclerosis in the ketosis-prone T2DM resembled those in the nonketotic T2DM 26,27 Wang et al 11 discovered that the prevalence of atherosclerosis was higher in the ketosis-prone T2DM group.This disparity may be due to the different study population.Second, ketosis-prone type 2 diabetes participants showed more severe bone destruction and more pronounced thyroid dysfunction.A similar conclusion was reached by Xu et al that ketosis or ketoacidosis may reduce the activity of osteoblasts, increase the activity of osteoclasts, and lead to β-CTX increase, while damaging the hypothalamic-pituitarythyroid axis, resulting in low T3 syndrome. 28

| Limitations
Several limitations in the current study should be addressed.First, because our center is a regional tertiary referral hospital with admissions of participants with relatively severe conditions, there may be a selection bias in the study population.But the homogeneous treatment process can help to reflect the characteristics of ketosisprone T2DM at the time of onset.Second, our study can only show the correlations between the risks and ketosisprone T2DM; it cannot explain causality.The underlying mechanisms need to be investigated.Third, a single cross-sectional design only reflects the characteristics of ketosis-prone T2DM participants at the time of onset and the long-term outcomes need further exploration.Fourth, the modeling and validation cohorts come from the same center, which implies that the results need to be further evaluated in a larger population.

| Future directions
Our study comprehensively showed the clinical characteristics of ketosis-prone T2DM and suggested the importance and necessity of identifying and managing this particular subtype.However, some related issues still need further clarification.What are the predisposing factors that contribute to the development of ketosis-prone T2DM?How about the pathogenesis?Which drugs are more suitable for the management of ketosis-prone T2DM?Are the long-term outcomes of ketosis-prone T2DM distinctive from other types of T2DM, such as islet function, complications, etc.? Therefore, longitudinal prospective cohort studies will be needed to explore these questions.

| CONCLUSION
In summary, ketosis-prone T2DM is a specific but common subtype in newly diagnosed T2DM participants in China, with specific clinical characteristics.Based on age, FCP, FFAs, HbA 1c , and urinary protein, our study provided an effective clinical risk prediction model for ketosis-prone T2DM, with an excellent discriminatory power of 0.806 while maintaining adequate calibration, which could help for precise classification and management.

AUTHOR CONTRIBUTIONS
Jia Zheng analyzed the data and wrote the paper.Yu Zhao and Xiaohong Wu contributed to study conceptualization.Shiyi Shen and Hanwen Xu researched the data.Ye Hu, Yubo Xing, and Yingxiang Song conducted data quality check.All the authors have approved the final version of article and were involved in the decision to submit the manuscript for publication.

F I G U R E 1
Flow chart of the enrolled patients in the modeling and validation cohorts.Participants in the modeling cohort were recruited from January 2015 to December 2016, and those in the validation cohort were recruited from January 2020 to December 2021.T2DM, type 2 diabetes mellitus.

F I G U R E 3
Nomogram for predicting ketosis-prone T2DM.The probability of ketosis-prone T2DM was determined by calculating the corresponding score for each risk factor.FCP, fasting C-peptide; FFAs, free fatty acids; HbA 1c , glycated hemoglobin A 1c ; T2DM, type 2 diabetes mellitus.F I G U R E 4The ROC curves of the model in the modeling and validation cohorts.(A) modeling cohort, (B) validation cohort.The AUC of the model in the modeling cohort was 0.806 (95% CI: 0.760-0.851),whereas it was 0.856 (95% CI: 0.803-0.908) in the validation cohort.AUC, area under the receiver operating characteristic curve; CI, confidence interval; ROC, receiver operating characteristic.

F
I G U R E 5 The calibration curves of the prediction model for ketosis-prone T2DM.(A) modeling cohort, (B) validation cohort.The calibration curves of modeling cohort and validation cohort both are close to a straight line with a slope of 1. T2DM, type 2 diabetes mellitus.
Clinical characteristics of the modeling and validation cohorts.