Artificial intelligence in health data analysis: The Darwinian evolution theory suggests an extremely simple and zero-cost large-scale screening tool for prediabetes and type 2 diabetes

Abstract

Aims

The effective identification of individuals with early dysglycemia status is key to reduce the incidence of type 2 diabetes. We develop and validate a novel zero-cost tool that significantly simplifies the screening of undiagnosed dysglycemia.

Methods

We use NHANES cross-sectional data over 10 years (2007–2016) to derive an equation that links non-laboratory exposure variables to the possible presence of undetected dysglycemia. For the first time, we adopt a novel artificial intelligence approach based on the Darwinian evolutionary theory to analyze health data. We collected data for 47 variables.

Results

Age and waist circumference are the only variables required to use the model. To identify undetected dysglycemia, we obtain an area under the curve (AUC) of 75.3%. Sensitivity and specificity are 0.65 and 0.73 by using the optimal threshold value determined from external validation data.

Conclusions

The use of uniquely two variables allows to obtain a zero-cost screening tool of analogous precision than that of more complex tools widely adopted in the literature. The newly developed tool has clinical use as it significantly simplifies the screening of dysglycemia. Furthermore, we suggest that the definition of an age-related waist circumference cut-off might help to improve existing diabetes risk factors.

Introduction

Risk-based screening tools for the identification of individuals with increased risk to develop diabetes have become very popular in the last decades [1], [2], [3], also thanks to the development of health data analysis methods based on artificial intelligence algorithms [4], [5], [6]. The early identification of high-risk individuals is a crucial task for public health administrations in order to reduce the incidence of several fatal and nonfatal complications among the population [1]. For example, it is known that individuals with prediabetes have enhanced future risk of cardiovascular disease [7], [8] and 3–12 times higher annual diabetes incidence than individuals with normoglycemia [9], [10], [11]. Furthermore, type 2 diabetes, which currently represents over 95% of diabetes cases and affects about 26.8 million people in the USA [12], [13], [14], is a major risk factor of death.

Recent studies have shown that the lifestyle modification or early pharmacological interventions can reduce the risk to develop dysglycemia, i.e. prediabetes or diabetes [15], [16], [17]. To this extent, clinical guidelines of several countries, including the USA [12], [18], often suggest to direct suitable treatments towards identified high-risk individuals based on models that involve the evaluation of risk factors. However, despite the accuracy reached by state-ot-the-art risk-based prediction models and the effectiveness of primary prevention approaches in reducing the risk of dysglycemia, in 2018, 88 million Americans (34.5% of the entire US population) were affected by prediabetes while 7.3 million (2.8%) had undiagnosed diabetes [14].

Identification of individuals with undetected dysglycemia is anything but a trivial task as it usually involves invasive and costly procedures [1], [19]. For this reason, alongside with models based on longitudinal cohort studies for the prediction of the future risk associated to diabetes [2], [20], tools capable to detect the current dysglycemia status of an individual based on the evaluation of zero-cost variables are highly desirable. They can serve as a pre-screening for more accurate diagnostic methods [12], [21]. To this end, detailed analyses of cross-sectional studies are required. In this framework, many recent works have been focused on obtaining dysglycemia screening tools for the US national population via variables that do not involve laboratory tests or that involve non-invasive laboratory examinations [4], [5], [6], [22], [23], [24], [25], [26], [27], [28], [29], [30], [31]. However, resulting tools are often unpractical for large-scale screenings as they exploit a large number of exposure variables connected by non-trivial relations or even not expressed by a mathematical equation.

To account for the need of a practical tool for the screening of dysglycemia in the USA population based exclusively on zero-cost non-laboratory variables, we performed an innovative analysis using US National Health and Nutrition Examination Surveys (NHANES) data from 2007 to 2016 through a unique artificial intelligence approach. The methodology used in our analysis is novel as it combines hybridization of state-of-the-art machine learning techniques such as genetic programming, which implements a scheme derived from the Darwinian evolutionary theory, and artificial neural networks [32], [33]. The combination of such peculiar techniques allows for the first time to perform a global and totally unbiased search for the optimal mathematical expression that connects non-laboratory variables, previously suggested as good dysglycemia predictors, with the undetected presence of prediabetes or undiagnosed type 2 diabetes.