Choosing the most appropriate existing type 2 diabetes risk assessment tool for use in the Philippines: a case-control study with an urban Filipino population

The prevalence of type 2 diabetes (T2DM) is rapidly increasing in low- to middle-income countries (LMICs) [1]. As T2DM’s prevalence increases in LMICs and it becomes a leading cause of mortality and morbidity, the health care and economic burden on affected nations also increases. The Philippines, considered an LMIC, had an overall diabetes prevalence of 6.0% as of 2013, and 1.7 million  people with undiagnosed diabetes, 3.2% of the adult population [2]. In the same year, there were 54,535 diabetes-related deaths in the country [2]. The Philippines is predicted to have over six million patients with diabetes by the year 2035 [1]. In terms of economic burden, the public Philippines Health Insurance Company alone paid out approximately 265 million Philippine pesos (approximately $5.2 million USD) for diabetes and diabetes-related conditions in 2006 [3]. This burden on health and social systems could be alleviated through the earlier detection of the risk of T2DM and prediabetes, thus potentially delaying or preventing the progression of the condition and decreasing potential complications [4].

The Philippines is in the early stages of implementing the World Health Organization’s Package of Essential Noncommunicable Disease Interventions for Primary Health Care, with some gaps in the implementation of the diabetes detection and treatment program. For example, though screening kits (glucometers and strips) and medications (oral hypoglycaemic) are available free of charge for the detection and treatment of diabetes [5], there is poor case detection for new cases of diabetes, no diabetes registry, and medications are often left unused in stockrooms of local health centres (Dr M.A. Mabolo, Philippine Department of Health, personal communication, June 16, 2014).

Timely diagnosis of T2DM could be facilitated by the use of validated diabetes risk assessment tools that can identify individuals who have a high risk of the disease, and clinicians could follow up to initiate diagnostic blood testing in those at risk. There is ample evidence that early treatment following early diagnosis, through lifestyle modifications and oral hypoglycaemic agents, could delay the progression of the disease [4].

Though many T2DM risk assessment tools have been developed worldwide, not every tool is as appropriate for the setting of the Philippines. Sensitivity and potential screening efficacy of other diabetes risk tools could be increased via adaptation and validation in a local setting [6]. Our study aims to evaluate six existing T2DM risk screening tools in the identification of individuals at risk of T2DM, when applied to an urban setting (Zamboanga City) in the Southern Philippines. The six tools evaluated were: the American Diabetes Association risk score (ADA) [7], the Canadian Diabetes Risk Score (CANRISK) [8], the Finnish Diabetes Risk Score (FINDRISC) [9], the Indian Diabetes Risk Score (IDRS) [10], an Indonesian Undiagnosed Diabetes Mellitus (UDDM) scoring system [11], and a proposed set of variables in elementary tool status from the Philippines [12]. These tools were selected based on one or more of the following criteria: tested for validity to predict diabetes, widely used clinically, and tested in an Asian population (preferably in a Filipino population). Though previous literature suggest that similarity in ethnicity does not necessarily affect the accuracy of the tools [13], using a similar methodology of applying these tools and comparing their results can elucidate which one will perform better in this population.

Of the 200 participants, 50 participants diagnosed with diabetes were recruited: 25 from a public diabetes clinic and 25 between three private diabetes clinics in Zamboanga City. One hundred and fifty participants not diagnosed with diabetes were recruited. All participants completed the study.

Combined T2DM and prediabetes prevalence in the undiagnosed population was 14% (21/150). Twelve (8%) had prediabetes while 9 (6%) had diabetes. The demographic data of the participants (Table 1) show that the majority of participants were female (n = 153; 76.5%). The mean age in the groups slightly varied with those with known diabetes being slightly older, and those with normal blood glucose the youngest. Educational attainment for the groups did not differ; most had an educational attainment of some high school or less (n = 66, 33.0%). The majority in each group were housewives, 32.5% overall. For those previously diagnosed with diabetes, the average number of years having had the disease was 5.5 years. For undiagnosed individuals, the mean fasting blood glucose for those identified with dysglycemia was 161.84 (SD = 86.42), and for those identified with normal blood glucose was 78.56 (SD = 9.5).

Performance of each existing risk assessment tool is shown in Table 2. Of the 71 participants in the study with dysglycemia, the FINDRISC correctly identified the most (n = 67). It had the highest sensitivity of the tools at 0.94, along with a specificity of 0.45. Conversely, the Filipino risk assessment tool correctly identified the most individuals who did not have dysglycemia, correctly identifying 94 of the 129 in that group. The Filipino tool had the highest specificity among the tools at 0.73, but also had the lowest sensitivity at 0.68.

Each risk assessment tool had a higher sensitivity and a lower specificity when used in this study than in their original studies. This comparison is shown in Table 3.

The Filipino tool had the highest positive predictive value (0.27) while the IDRS and the Indonesian UDDM screening tool had the highest negative predictive values, both at 0.96. The Filipino tool had the highest LR+ of 2.49. All the remaining risk screening tools had a LR+ value of less than 2. The risk screening tool with the lowest LR- was the FINDRISC with a LR- of 0.13. The Filipino tool had the highest LR- of 0.45, while the other tools’ LR- ranged from 0.2 to 0.4.

The tool with the largest AUC was the FINDRISC, AUC = 0.80. Overall performance of each risk assessment tool for ROC curves and AUC, compared to values in its original publication, are shown in Table 4 and Fig 1; the FINDRISC performed the best.

The prevalence rate of undiagnosed dysglycemia (diabetes or prediabetes) in this population was 14%, which was comparable to a previous study conducted in the Philippines, where the overall prevalence of dysglycemia was 14.2% (7.2% diabetes and 7.0% prediabetes) [16]. In our study population, we discovered that the FINDRISC was the most effective tool, having the highest sensitivity (0.94) and AUC (0.80), and a high negative predictive value (0.95). These scores are balanced by a low specificity (0.45). On the other end, the Filipino tool had the highest specificity (0.73), but the lowest sensitivity (0.68). In the case for screening for T2DM, clinicians and public health doctors are more concerned about the sensitivity of the test. By increasing the sensitivity of a screening test, false negatives are decreased, and therefore fewer cases are missed. However, specificity should not be compromised by doing so, as false positives can create unnecessary anxiety for those screened, and it may also include unnecessary expenditures for the individual or the health care system.

The FINDRISC’s high sensitivity and lower specificity indicates that although most participants would have been correctly classified as high risk, a considerable portion (55%) would have been subjected to unnecessary follow-up as they would have been classified as high risk despite an absence of the condition. However, this would be a better option compared to the Filipino tool, that has a high specificity but the sensitivity would indicate that a considerable number of respondents (32%) may be missed.

The CANRISK demonstrated the second highest specificity (0.54) though it was lower than the FINDRISC in its sensitivity (0.86). Though the CANRISK was derived from the the FINDRISC, it was adopted to fit the multi-ethnic Canadian population. The Philippines has a more homogenous population, so the Canadian tool with its focus on a multi-ethnic population is likely not the best fit in this population. This may explain why the FINDRISC showed better predictive results.

Compared to original publications, the positive predictive values of FINDRISC, ADA, and IDRS were higher in this study. This could be attributed to the higher prevalence of the disease in our study population. In general, highly specific and sensitive tests will have a high positive predictive value in a population with a high prevalence of the disorder [4].

Statistically, the larger the area under the ROC curve (AUC), the more the accurate a tool can be considered in its overall performance. An AUC of 0.9–1.0 is considered excellent, 0.8–0.9 very good, 0.7–0.8 good, 0.6–0.7 sufficient, 0.5–0.6 bad, and less than 0.5 considered not useful [18]. While the FINDRISC had the highest AUC in this study (0.80), it was lower than the AUC from the original validation study for the tool. This could be a result of the differences in the participant characteristics, such as differences in age (> 25 years old in the original study, > 18 years old in the current study) and ethnicity (primarily white versus a homogenous Filipino population). The AUC in this study also varied from a previous Philippines-set study using the FINDRISC, which had a lower AUC of 0.63 [14]. However, overall, the performance of the FINDRISC in both studies showed sufficient diagnostic accuracy. Conversely, the ADA and the Indonesian UDDM scoring system had a higher AUC in the current study compared to their original publications. This could be due to the question of history of gestational diabetes being included in this study for the ADA, while it was not in its original study. For the Indonesian UDDM scoring system, since the variables of hypertension, central obesity, and obesity were not defined, cut-off levels used for this study were adopted from the WHO, which probably differed from the cut-offs used in the original publication.

Finally, this study found that an adjustment in the cut-off score of our chosen T2DM risk assessment tool, the FINDRISC, could improve optimal sensitivity and specificity. A 2013 study testing the effectiveness of the FINDRISC in the Filipino population found that the tool was effective at screening undiagnosed diabetes (AUC = 0.738), and had the highest diagnostic accuracy when the cut-off score distinguishing those at risk of diabetes was raised from 7 to 9 [15]. Another study in the Filipino population showed that with a cut-off score > 9, the FINDRISC had moderate diagnostic accuracy (with a sensitivity of 74.1%, specificity of 52.6%, and AUC of 0.63) [14]. In our recalibration, a cut-off score of 11 produced the most optimal sensitivity (0.86) and specificity (0.73).

Within already existing risk assessment tools, this study identified the FINDRISC as the most accurate diabetes risk assessment tool for the Philippines, based on its high sensitivity (0.94), negative predictive value (0.95), and AUC (0.80). Our data show increased accuracy in the use of the FINDRISC when the cut-off score was raised from 7 to 11, resulting in a better balance between sensitivity (0.86) and specificity (0.71). The accuracy of the FINDRISC and its modifications in the Philippines suggest that it is appropriate for initial population screening.