Diabetes risk score in Qatar: Model development, validation, and external validation of several models.

Aim: To establish a simple risk score for identifying individuals at high risk of developing Type 2 Diabetes Mellitus (T2DM) in Qatar, based on easy-to-obtain variables. Materials and Methods: A cross sectional sample of 2000 individuals from Qatar BioBank, obtained for the development cohort, was evaluated using the Least Absolute Shrinkage and Selection Operator (LASSO) for feature selection along with logistic regression to determine the predictive variables for the risk of T2DM along with impaired glucose metabolism (IGM). Complex machine learning analyses were performed for comparative purposes. Another sample of 1000 participants was subsequently obtained for external validation of the developed models. Several existing scoring models screening for T2DM were evaluated and compared to the model proposed by this study. Results: 1660 participants were included in the analysis which showed that age, gender, waist-to-hip-ratio, history of hypertension (HTN), history of hypercholesteremia (HCL) and levels of educational were statistically associated with the risk of T2DM and constituted the Qatari diabetes risk score (QDRISK). In addition to the 6 aforementioned variables constituting the QDRISK, the IGM model showed that BMI was statistically significant. The QDRISK performed well with an area under the curve (AUC) 0.870 (95%CI 0.843, 0.896) in the development and 0.815 (95% CI: 0.765, 0.864) and external validation cohort. Compared to the majority of evaluated scores, the QDRISK showed overall better accuracy and calibration. The IGM model performed moderately high with AUC 0.796 (95% CI: 0.774, 0.819) in the development cohort and 0.774 (95% CI: 0.740, 0.809) in the external validation cohort. Conclusions: Our study investigated risk factors for developing T2DM as well as IGM and developed Qatari-specific risk scores based on demographic and anthropometric factors to identify high risk individuals. This simple, cost-effective and convenient tool can guide the development of a nationwide primary prevention program in Qatar.


Introduction:
Diabetes is one of the most prevalent chronic diseases worldwide. The International Diabetes Federation (IDF) estimated that approximately 463 million individuals aged between 20 -79 are diagnosed with diabetes in 2019, and that this figure is expected to rise to 700 million by 2045 [1].
Qatar is reported as one of the most affected populations in the world with an age-adjusted prevalence of 15.6% [1]. Along with its high prevalence, diabetes contributes to a number of major complications including cardiovascular diseases, nephropathies, retinopathies, and neuropathies. [2]. Further, diabetes imposes a considerable economic burden, reported as $327.2 billion in the U.S in 2017 [3] In addition, the most recent estimate of undiagnosed cases of diabetes was as high as 231.9 million. Consequently, early detection of such individuals might prevent or delay the onset of Type 2 Diabetes Mellitus (T2DM) as it contributes to 90% of diabetes [1]. Several risk factors were associated with increased incidence of T2DM, including; age, ethnicity, family history, obesity, nutrition, smoking, and physical inactivity [1,2]. Screening for T2DM can be conducted with either non-invasive risk scores or invasive laboratory measure of glycosylated hemoglobin (HBA1c), fasting plasma glucose, or random blood glucose (RBG) levels.
In comparison to invasive laboratory testing, screening models are both time and cost effective, along with its convenient applicability in clinical settings [4]. Several screening risk scores for T2DM were developed for different populations [5][6][7][8][9][10][11][12][13][14][15][16][17][18]. Awad et al. [19] recentely developed a screening tool for T2DM in Qatari using data simulation of population characteristics and disease prevalence. However, screening models need to be developed from real population data to reflect the true nature of disease burden and account for confounding factors. To the best of our knowledge, there is no study reporting on the development of a risk score for T2DM in Qatar.
Therefore, this study aims to generate a simple and reliable screening score, that is tailored specifically for the Qatari population and can be implemented within primary healthcare settings or by individuals in the community to identify ones at high risk of developing T2DM in Qatar.

Study design, setting and data source
This cross-sectional study was conducted based on data provided from the Qatar BioBank (QBB).
QBB is a population-based project which recruits adults who are either nationals or long-term . CC-BY 4.0 International license It is made available under a is the author/funder, who has granted medRxiv a license to display the preprint in perpetuity.

(which was not certified by peer review)
The copyright holder for this preprint this version posted April 7, 2021. ; https://doi.org/10.1101/2021.04.04.21254900 doi: medRxiv preprint residents (≥ 15 years) in Qatar aged 18 to 89 years, and follows them up to measure a wide range of various health parameters. Recruitment of subjects started in 2012 and obtained information of more than 17,065 participants. Participation is voluntary, based on personal recommendations of friends and family, via social media, bookings via the website (available at: www.qatarbiobank.org.qa), or through phone calls.
Upon obtaining written informed consent from participants in their first visit to the QBB facility at Hamad Medical City Building 17, Doha, Qatar, they undergo a 5-stage interview, along with physical and clinical measurement sequence, the interview averaging 3 hours. Comprehensive self-reported questionnaires are completed on health behaviors, past medical history, lifestyle variables, physical activity, mental well-being, environmental factors' exposure and other parameters, followed by physical examination to obtain anthropometric measurements, blood pressure, electrocardiogram, bone density and other measures by trained research personnel.
Afterwards, enrolled participants provide saliva, blood and urine samples, which undergo analysis at Hamad General Hospital Medical Centre Laboratory in Doha, Qatar [20,21].
Participants are contacted via e-mail, messages, or phone calls to present for a follow up visit at QBB facility every 5 years, where they undergo the same procedures as the baseline visit. QBB is regularly recruiting further participants, aiming to represent the population at Qatar by reaching 60,000 study participants. For detailed description of procedures of biological marker assessments please review Qatar Biobank for Medical Research [22]. Ethical approval for the study was obtained from the QBB Institutional Review Board (Ex-2018-RES-ACC-0097-0043).

Inclusion and exclusion criteria
A cross sectional study was carried out involving participants provided by QBB aging 18 years or older. Candidates with a history of type 1 diabetes mellitus or gestational diabetes were excluded from this study. From the remaining participants, only those with HbA1c and random plasma glucose (RPG) levels available were included in the analysis.

The Qatari diabetes risk score (QDRSIK)
Participants with HbA1c ≥6.5% and/or history of diabetes and/or history of anti-diabetic medication were considered diabetic individuals based on this proxy composite criterion. All other participants were deemed non-diabetic.
. CC-BY 4.0 International license It is made available under a is the author/funder, who has granted medRxiv a license to display the preprint in perpetuity. (which was not certified by peer review) The copyright holder for this preprint this version posted April 7, 2021. ; https://doi.org/10.1101/2021.04.04.21254900 doi: medRxiv preprint

Screening for impaired glucose metabolism
For comparative analysis of performance measures, we developed a second model screening for Impaired Glucose Metabolism (IGM), where participants with HbA1c level of ≥5.7% [2] and/or RPG levels ≥140 mg/dL (7.8 mmol) [23], and/or history of diabetes and/or history of anti-diabetic medication were deemed having IGM, while the remaining participants were labeled non-IGM.

Statistical analysis
Differences between categorical variables were compared using Chi-square test and Fisher's Exact Test and were reported as frequencies and percentages, while differences between continuous variables was examined using Mann-Whitney test and were reported as median (IQR).
To aid with clinical decision-making, continuous variables were dummy coded into (0 vs. 1) representing above and below cut-offs pre-determined. While this approach is often criticized, we believe it is practical and more applicable in primary healthcare settings.

Least Absolute Shrinkage and Selection Operator
To avoid over-fitting, ten variables including: age, gender, history of hypertension (HTN), hypercholesterolemia (HCL), waist-to-hip ratio (≥1 for males and ≥0.86 for females), hours of sleep, BMI (normal, overweight, and obese) [24], educational level, vegetable and fruit consumption were entered into the Least Absolute Shrinkage and Selection Operator (LASSO) binary logistic regression against the two composite binary outcomes detailed above.
The L1-penalized LASSO was utilized for the multivariable analysis with 10-fold cross validation for internal validation. LASSO is a machine-learning logistic regression which penalizes the size of the coefficients of based on the value of the hyperparameter lambda (λ). Coefficients of weaker variables are reduced towards zero with larger penalties, while only the strongest features remain in the model. Feature selection was carried out by the minimum value of lambda (λ.min). The R package "glmnet" was used to conduct the LASSO regression. Afterwards, variables recognized by the regression model entered a logistic regression model.

Complex Machine Learning Analysis
The second strategy used in this study one relied on further sophisticated machine learning (ML) analysis using the final significant variables in the logistic regression model to build four ML . CC-BY 4.0 International license It is made available under a is the author/funder, who has granted medRxiv a license to display the preprint in perpetuity. (which was not certified by peer review) The copyright holder for this preprint this version posted April 7, 2021. ; https://doi.org/10.1101/2021.04.04.21254900 doi: medRxiv preprint models, using random forest (RF), gradient boosting machine (GBM), XgBoost as well as deep learning (DL). These models also used 10-fold cross-validation. Statistical analysis for the aforementioned ML models was performed using R package "h2o" (version 3.32.1.1).

External validation of several diabetes risk scores
We attempted to externally validate diabetes risk score where data for their categories was available. The choice of these models was either based on practicality and ease of use. These included the Danish [25], Thai [17], and the Data from the Epidemiological Study on the Insulin

Assessment of accuracy and calibration
The accuracy of QDRSIK was assessed using the area under the receiver-operator characteristic curve (AUC) of the logistic and ML models and compared them to those of the six aforementionedmentioned models. Calibration of the models was assessed using the "rms" package in R with  CC-BY 4.0 International license It is made available under a is the author/funder, who has granted medRxiv a license to display the preprint in perpetuity.
Compared with the Non-diabetics, patients with T2DM participants were more likely to be older, obese, with history of HTN and HCL, and with lower educational levels (Table 1). In the second model, participants were categorized into those with IGM (n = 606, 32.6%) and non-IGM (n = 1254, 67,4%). Those with IGM were more likely to be older, obese, with history of HTN and HCL (Table 1). Subjects in the external validation cohort had similar overall characteristics compared to the development cohort (Table 2).

Logistic regression
The LASSO analysis excluded the following variables: BMI, vegetable and fruit consumption ( Figure S1-A & S1-B). Subsequently, the remaining 7 variables (age, gender, waist-to-hip-ratio, history of HTN, history of HCL, educational levels and hours of sleep) entered a logistic regression (LR) analysis and 6 of them were found to be significant predictors of T2DM. Beta-coefficients of significant variables were multiplied by 10 and rounded to the nearest integer and therefore constructed the risk score. These included age ≥55 (26 points), 36-54 (12 points), history of HCL (12 points), male gender (9 points), abnormal Waist-to-hip-ratio (7 points), history HTN (7 points), education category 2 (2 points), category 3 (8 points) (please see Table 1). Although vegetable and fruit consumption (< once daily) were excluded by the LASSO, they were allocated 1 point  Table S1. Accordingly, subjects with 0-31, 32-42, >47 scores were deemed as at low, moderate, and high risk of developing T2DM, respectively.
The model showed overall good calibration curves in both cohorts ( Figure S2).

Comparative analysis with complex machine learning models
. CC-BY 4.0 International license It is made available under a is the author/funder, who has granted medRxiv a license to display the preprint in perpetuity.

(which was not certified by peer review)
The copyright holder for this preprint this version posted April 7, 2021. ; https://doi.org/10.1101/2021.04.04.21254900 doi: medRxiv preprint The performance of the ML models in the validation cohort was comparable to that of the regression model ( Figure 1-A & 1-B) with DL AUC 0.817, GBM AUC 0.7452, RF AUC 0.814, and XGB AUC 0.717 (Figure 1-C). Sensitivities, specificities, PPVs and NPVs can be found in Table   S2.

Screening for impaired glucose metabolism
The LASSO feature-selection excluded only vegetable and fruit consumption ( Figure   and Qatari mathematical model [19]; generated using the Qatari population characteristics and estimates of various disease prevalence; they all showed lower performance compared to the QDRISK with AUCs ranging (0.733-0.796) ( Figure 2) and comparable calibration curves and other validation metrics ( Figure S5 and Table S5). Interestingly, performance of the Danish score  (Figure 1-A & 1-B), respectively, along with inferior calibration curves compared to the rest of the models. Therefore, we did not assess their validation metrics. Comparison of validation of the QDRSIK and externally validated models is shown in Table S5.

External validation of diabetes risk scores
. CC-BY 4.0 International license It is made available under a is the author/funder, who has granted medRxiv a license to display the preprint in perpetuity.

Discussion
This study presents the first screening tool tailored specifically for the Qatari population, that identifies individuals at high risk of developing T2DM. The developed tool is based on 8 risk variables including; age, gender, waist-to-hip-ratio, history of HTN, history of HCL, educational levels, along with vegetable and fruit consumption. The proposed score can be utilized by healthcare professionals, or individuals in the community as the risk factors are easy to obtain and measure.
Age was the highest predictive variable for the risk of T2DM in our score. This finding is consistent with other established scores [6,7,9,10,14,18]. However, age was non-predictive in the score by Wilson et al. [12]. Several studies proposed that the risk of developing T2DM increases with older ages [6-10, 13-15, 33]. The male gender was also an important factor in our model in parallel with the studies by Aekplakorn, Wang and Kahn et al. [13,15,17].

(which was not certified by peer review)
The copyright holder for this preprint this version posted April 7, 2021. ; https://doi.org/10.1101/2021.04.04.21254900 doi: medRxiv preprint Compared to an estimated $77,445 10-year medical expenditures attributed to a patient with diabetes [3,35], the 10-year costs of lifestyle interventions per capita were $4,601 in the DPP and its Outcome study [36]. This highlights the substantial cost-saving capability of lifestyle changes.
Notwithstanding its potential public health implications, two limitations to this study are important to highlight. First, we could not perform multiple imputation on family history of diabetes and HTN due to missing data (> 20%). Lack of this variable in our score might be a limitation, since it reflects the genetic predisposition for diabetes and is a significant factor that could increase the risk for T2DM [37, 38]. Nonetheless, it was not included in the widely implemented Finnish risk score, and the accurate German score [5,8]. Finally, the use of cohort design could have improved the generalizability and reliability of the findings. On the other hand, our study was specifically designed and validated for the Qatari population, along with including an adequate sample size with a wide age range. The key advantages of this tool are its cost-effectiveness, practicality, and relying on factors that are not difficult to obtain. Hence, we recommend applying it at the level of the population as a screening tool for identification of high-risk individuals who would benefit from lifestyle interventions. Further studies should externally validate this risk score in larger cohorts within the Qatari population in the future.
In conclusion, our study investigated risk factors and developed a validated risk score designed for the Qatari population to assess the risk of developing T2DM as well as IGM based on demographic and anthropometric factors. It demonstrates a simple, cost-effective and convenient tool which may be used to identify individuals at high risk. Such individuals should undergo further blood testing and early lifestyle modifications aiming to decrease the incidence of T2DM and IGM. Further studies are needed to validate the utility of the present risk scores in larger cohorts in Qatar and externally.

Conflicts of Interest
The authors declare that there is no conflict of interest regarding the publication of this paper.

Funding Statement
This research did not receive any grant from funding agencies in the public, commercial, or notfor-profit sectors.
. CC-BY 4.0 International license It is made available under a is the author/funder, who has granted medRxiv a license to display the preprint in perpetuity.

Acknowledgments
We would like to thank Qatar Biobank project for providing the data.
. CC-BY 4.0 International license It is made available under a is the author/funder, who has granted medRxiv a license to display the preprint in perpetuity.

(which was not certified by peer review)
The copyright holder for this preprint this version posted April 7, 2021. . CC-BY 4.0 International license It is made available under a is the author/funder, who has granted medRxiv a license to display the preprint in perpetuity.

(which was not certified by peer review)
The copyright holder for this preprint this version posted April 7, 2021. ; https://doi.org/10.1101/2021.04.04.21254900 doi: medRxiv preprint . CC-BY 4.0 International license It is made available under a is the author/funder, who has granted medRxiv a license to display the preprint in perpetuity.

(which was not certified by peer review)
The copyright holder for this preprint this version posted April 7, 2021. . CC-BY 4.0 International license It is made available under a is the author/funder, who has granted medRxiv a license to display the preprint in perpetuity.

(which was not certified by peer review)
The copyright holder for this preprint this version posted April 7, 2021. ; https://doi.org/10.1101/2021.04.04.21254900 doi: medRxiv preprint . CC-BY 4.0 International license It is made available under a is the author/funder, who has granted medRxiv a license to display the preprint in perpetuity.
. CC-BY 4.0 International license It is made available under a is the author/funder, who has granted medRxiv a license to display the preprint in perpetuity. (which was not certified by peer review) The copyright holder for this preprint this version posted April 7, 2021. ; https://doi.org/10.1101/2021.04.04.21254900 doi: medRxiv preprint is the author/funder, who has granted medRxiv a license to display the preprint in perpetuity. (which was not certified by peer review) The copyright holder for this preprint this version posted April 7, 2021. ; https://doi.org/10.1101/2021.04.04.21254900 doi: medRxiv preprint  is the author/funder, who has granted medRxiv a license to display the preprint in perpetuity. (which was not certified by peer review) The copyright holder for this preprint this version posted April 7, 2021. ; https://doi.org/10.1101/2021.04.04.21254900 doi: medRxiv preprint