Prediction Models for Early Childhood Obesity: Applicability and Existing Issues

Clinical prediction models are becoming increasingly prevalent in the medical literature [1]. These models be can either diagnostic or prognostic. Diagnostic models aim to estimate the likelihood of an individual currently having a disease/illness, whereas prognostic models aim to estimate the risk of an individual experiencing a particular outcome at a specified future time. Clinical prediction models are commonly developed using regression modelling, but other approaches such as machine-learning techniques and decision trees have also been utilized [1-4].

Prediction models exist for a wide variety of health outcomes, such as cardiovascular disease [5], lung cancer [6], acute kidney injury [7], and infertility treatment success [8]. There is an overwhelming number of models for the prediction of specific outcomes, with, for example, more than 350 prediction models developed for cardiovascular disease in the general population [5]. Conversely, very few studies have focused on the prediction of overweight or obesity in early childhood (between 2 and 8 years of age) [9-16], which is the focus of this review.

The prevalence of childhood obesity increased in every region of the world between 1975 and 2016 [17]. It is now estimated that, worldwide, 41 million children under the age of 5 years are overweight or obese [18]. In New Zealand, one third of 4- to 5-year-olds are overweight or obese [19]. Childhood obesity is associated with both long-term and immediate health risks. In the long term, it is predictive of diabetes risk [20, 21], and it is associated with premature mortality [22] and increased cardiovascular risk [21, 22], predominantly as a result of a substantially greater risk of adult obesity [20, 23-26]. In the short term, childhood overweight and obesity are associated with psychosocial distress [22, 27-29], a lower health-related quality of life [30], and a range of physical comorbidities [29, 31]. For example, even among preschool children (aged 3–6 years), an increase in body mass index (BMI) of 1 kg/m² has been associated with increased systolic and diastolic blood pressure [32]. Furthermore, conditions that were previously considered restricted to adults (e.g., type 2 diabetes and sleep apnea) are now being diagnosed more and more frequently in obese youth [23, 29, 31].

Prevention, rather than treatment, has been argued to be more successful in reducing childhood obesity [23], as obesity is difficult to reverse once established [23, 33]. Interventions to reduce obesity in adults have mostly focused on lifestyle changes alone, and they have struggled to maintain results 36 months after intervention [34]. Weight loss medication and bariatric surgery are associated with side effects and risks [34, 35]. In addition, as childhood obesity is associated with health risks even in very young children [32], it would be preferable to prevent obesity from occurring in the first place.

There are a number of potentially modifiable risk factors that are known to be associated with early childhood obesity, such as excessive gestational weight gain [36, 37], maternal pre-pregnancy BMI [36-38], maternal smoking during pregnancy [36-38], and rapid infant weight gain [36, 38-40]. However, despite the identification of these potentially modifiable risk factors, reviews of interventions to prevent early childhood obesity have reported inconsistent results [41-44].

To our knowledge, 7 studies have been published reporting models for predicting overweight/obesity in young children aged ≤8 years [9-15]. Two studies reporting the prediction of childhood obesity and/or overweight beyond 8 years of age were excluded [40, 45]. The study by Morandi et al. [13] reported a number of models with the outcome at various ages, and only those falling within our age range criteria were considered. Additionally, we excluded certain models from 2 studies that predicted obesity and/or overweight within approximately 12 months from the time of prediction [11, 15], but their other models were included in this review.

The details of the derived prediction models included in this review are examined in Table 1, while Table 2 outlines the respective validation models. These models are all prognostic, aiming to calculate the risk of an infant or toddler becoming overweight or obese at a future time point. Six of these studies used regression modelling [9, 10, 12-15], while the remaining study used data-mining methods [11]. The models used a variety of modifiable predictors, including maternal prepregnancy BMI [10, 12-15], paternal BMI [10, 12, 13], and maternal smoking during pregnancy [10, 13] (Table 3; online suppl. Table 1; see www.karger.com/doi/10.1159/000496563 for all online suppl. material). Conversely, non-modifiable risk predictors included infant birth weight [10, 12-14], infant weight gain [10, 11], and sex [10-12, 15] (Table 3; online suppl. Table 1). Table 3 outlines the various predictor variables used in the derivation models, as well as their respective OR where available. Additionally, online supplementary Table 1 describes the predictor variables used in each derivation model. In general, the models focused on predictor variables that can be routinely collected in clinical practice, although 3 models also included paternal BMI [10, 12, 13], and that may not be consistently available. While external validation is strongly recommended following development and internal validation of a prediction model [2, 46], this was only done for 3 of the 7 models [13, 15, 47].

The area under the receiver operating characteristic curve (AUROC) is a useful indicator of a model’s ability to discriminate between those at high risk and those at lower risk of a given outcome [48]. The following are generally accepted interpretations of AUROC values: poor (< 0.60), possibly helpful (≥0.60 but < 0.70), acceptable (≥0.70 but < 0.80), excellent (≥0.80 but < 0.90), and outstanding (≥0.90) [48, 49].

None of the models reported here had a poor AUROC, although they did vary considerably (Table 1). The model with the lowest AUROC (i.e., 0.67) predicted overweight/obesity at 7 years, but the authors did not validate it, as they did not deem it clinically useful [13] (Table 1). The highest reported AUROC (i.e., 0.87) was for a model to predict overweight at 2 years based on data gathered at 4.5–6.5 months [15] (Table 1). It is possible that the short time between the prediction age and the outcome age partly explains this high AUROC value. In general, regression models that were supplemented by data on infant weight gain [14, 15] produced higher AUROC values than those that used birth data alone [12, 13], with one exception [10]. These findings suggest that the inclusion of infancy weight gain could increase a model’s discriminative ability. However, this would occur at the cost of delaying the model prediction and consequently any intervention, which might affect potential success.

Sensitivity and specificity are key determinants of a model’s predictive ability. Sensitivity refers to a model’s ability to correctly predict the individuals who either have or will develop the outcome of interest [1]. Specificity refers to a model’s ability to correctly rule out the individuals who do not have or will not develop the condition of interest [1]. The ideal model would produce 100% sensitivity and 100% specificity, but this is considered to be unrealistic [48].

The choice of where to set the model’s risk threshold (i.e., the level of risk at or above which an individual will be classified as having an outcome) will impact on the sensitivity and specificity levels produced by the model [1]. Some models (e.g., the Framingham risk prediction tool for cardiovascular disease) use thresholds that are supported by clinical guidelines [50], but there are no similar guidelines for the prediction of early childhood obesity [16]. Indeed, some researchers have reported their results at a variety of risk thresholds without any discussion as to the most appropriate one to use [16]. In this context, the model reported by Robson et al. [14] illustrates the impact of varying risk thresholds can have on sensitivity and specificity. At the 25th percentile risk threshold, their model produced 95% sensitivity and 35% specificity, while at the 90th percentile the sensitivity was 30% and the specificity 99% [14]. Decision curve analysis, relative utility curves, and net benefits are methods for determining both risk thresholds and the clinical usefulness of prediction models [51-53]. However, to date, none of these methods have been utilized in the development or evaluation of early childhood obesity prediction models.

Importantly, the characteristics of the proposed intervention and the prevalence of obesity in the target population should also influence the choice of risk threshold. A model that gives a high sensitivity but a low specificity will potentially identify most, if not all, infants likely to develop obesity, but it will also mistakenly identify lots of infants as being at risk when they are not (i.e., a high false-positive rate). Such a scenario may be considered acceptable for early childhood obesity interventions that encourage healthy lifestyle behaviors that would not be physically harmful to participants, such as breastfeeding, which is widely accepted as beneficial for the health of infants and mothers alike [16, 54]. On the other hand, such models (i.e., with a high sensitivity and a low specificity) would generally be no different from non-targeted community-wide interventions, yet parents may experience avoidable stigmatization if they are unnecessarily invited to participate in an intervention based on the assumption that their infant will become an obese child. When interviewed about the use of interventions resulting from early childhood obesity risk prediction models, parents were receptive to the concept, but also expressed fears of judgement and concerns about feeling upset, ashamed, and guilty [55]. However, a pilot study of lifestyle counselling to prevent early childhood obesity showed that the intervention was generally acceptable [56], suggesting that any anxiety may be a function of the initial prediction rather than the intervention itself.

Conversely, while a model that produces a low sensitivity and a high specificity will be unlikely to misclassify infants as being at risk of obesity, it will likely fail to identify a large number of infants who will eventually become obese children. This would mean that, although fewer families are at risk of unnecessary stigmatization, more infants would potentially go on to experience known physical and psychosocial obesity comorbidities, including psychological distress due to weight-related stigmatization [20-22, 27-32].

The availability of resources for intervention should also influence the choice of risk threshold. If resources are scarce, a higher risk threshold may be preferable, thereby ensuring that those at a greater risk of developing obesity are prioritized. Alternatively, a low-cost and easily resourced intervention may warrant a lower risk threshold [16]. Lastly, the prevalence of obesity in the target population should also be considered when setting the risk threshold. For populations with a high prevalence of obesity, prediction models would be more useful to identify those at the highest level of obesity risk. On the other hand, in populations where the prevalence of obesity is low, prediction models would likely be more useful, aiming to identify infants across a wider range of the obesity risk spectrum.

Prediction models are primarily developed to improve healthcare-related decision making and therefore patient health [57]. In the case of early childhood obesity models, their purpose is to predict an infant’s likelihood of being obese by a specified future point that is still early in the individual’s life so that preventative measures can be taken to maximize long-term health benefits. As prediction models invariably have a less-than-perfect performance (i.e., sensitivity and specificity levels below 100%), Moons et al. [57] suggested that only impact analysis studies can truly determine their true clinical value. Such studies are generally developed using a comparative intervention approach to assess health outcomes obtained by using the prediction model versus not using it (i.e., standard care) [2]. Although all early childhood obesity prediction models display a less-than-perfect performance, as of yet, none have been subjected to an impact analysis study.

Nonetheless, 2 models have been developed into tools for use by clinicians or parents [16]. Firstly, Santorelli et al. [15] developed their model into a smartphone application called “Healthy Infant Weight?”. The application allowed parents to obtain their baby’s risk of obesity at 2 years by entering their baby’s sex, birthweight, and weight change between birth and 6, 9, or 12 months. Adding maternal BMI was optional but, as can be seen in Table 2, this made little difference to the sensitivity or specificity of the model [15]. Unfortunately, this mobile application was withdrawn due to funding issues [Santorelli, pers. commun.] and no research has been published regarding any outcomes achieved through its use [16].

The second tool is the Proactive Assessment of Obesity Risk during Infancy (ProAsk) [58], which was developed using the prediction model derived by Weng et al. [10] and later externally validated by Redsell et al. [47]. It should be noted that the name of the tool is somewhat misleading given that it does not actually predict the risk of obesity in particular, but instead the risk of overweight. It is also unclear what outcome age was used by ProAsk, although the development and validation models used ages 3 and 5 years, respectively [10, 47, 58]. ProAsk was designed to be used on tablets and adopted the following predictors: birth weight and length, infant’s current weight, maternal and paternal weight, maternal smoking during pregnancy, and breastfeeding. The tool also incorporated a therapeutic wheel of suggested behavioral changes to reduce the infant’s risk [58]. A feasibility study into the use of ProAsk by UK public health nurses with parents of infants aged 3 months produced disappointing results: the recruitment target was not achieved and attrition rates were high [58]. There were also methodological issues identified, such as public health nurses not fully adhering to the study protocols. On a positive note, 8 out of 12 parents interviewed at follow-up found ProAsk to be engaging. However, on balance, the researchers concluded that it would require significant additional resources for the delivery of ProAsk to be a feasible part of the public health nurses’ role [58].

Both of the above mentioned practical applications of early childhood obesity prediction models utilized digital technology. The relentless increase in worldwide smartphone ownership rates [59] makes them attractive tools for delivering public health interventions. A review of 23 studies on mobile phone applications aimed at changing health behaviors found that 19 reported statistically significant improvements in the behavior of interest [60]. Applications that were more effective utilized behavior change theory, were user friendly, had personalized features or feedback, and had involvement of health professionals [60]. These findings were echoed by Litterbach et al. [61], who found that participant engagement with their infant feeding application was increased by tailored content, user friendliness, and credible information sources. Breastfeeding mothers reported more confidence to continue breastfeeding, while formula-feeding mothers reported more confidence to feed in response to infant hunger cues [61]. These findings suggest that there may be potential for mobile phone applications based on early childhood obesity prediction models (such as Healthy Infant Weight? [15]) to be successfully implemented. However, it should also be noted that sole reliance on smartphone applications for early childhood obesity interventions may further alienate populations that are already traditionally difficult to engage. Worldwide, people with lower incomes and less education report less smartphone ownership and Internet access [59].

The development of tools such as ProAsk [58], which was designed to be used by health professionals on handheld devices, may serve to increase engagement with difficult-to-reach populations as they are easily portable. Indeed, despite the overall disappointing results, 33% of the participants recruited by the ProAsk study were from areas of increased social deprivation. In addition, Salvy et al. [62] identified the US home visitation program for at-risk families as an ideal opportunity for interventions to prevent early childhood obesity, as these programs are already engaging with high-risk families in the home environment where the obesity-promoting behaviors are likely to occur. Thus, it may be that future research could learn from the methodological issues identified in the ProAsk feasibility study [58] and successfully utilize risk prediction models for early childhood obesity prevention in the home setting. However, without any kind of impact analysis study, it is not possible to determine the true clinical value of any early childhood obesity risk prediction model, irrespectively of the use or not of digital technology. This may not be possible until clinically relevant risk thresholds and sensitivity and specificity levels have been agreed upon for early childhood obesity prediction [16].

The rationale for obesity prediction models is to identify young children at risk of obesity for whom targeted interventions can be implemented. In addition, these prediction models could be even more useful if they contained modifiable factors that relate to the preconception and pregnancy phases. This means that early interventions could be put in place to improve the outcome for children born from subsequent pregnancies within a given family. Further, the focus of interventions could also be extended to young women who may become pregnant, to ensure the best possible weight-related outcomes for their offspring. Healthcare professionals communicating the risk of early childhood obesity to families may need specific training regarding the discussion of a potentially distressing topic, and the provision of accurate information about lifestyle changes that can reduce the risk.

To date, 7 studies have been published reporting early childhood obesity prediction models [9-15]. The AUROC, sensitivity, and specificity produced by the models varied greatly, as did the risk thresholds applied to them. Risk thresholds in particular should be defined after careful weighing of the consequences of false-positive versus false-negative predictions in the context of existing (if any) clinically relevant guidelines [16]. Although there are no such guidelines for obesity prediction, methods to assist with the identification of appropriate risk thresholds do exist. However, these have not been utilized in the development of any of the previously published models for the prediction of early childhood obesity. Impact analysis studies of these models are currently absent from the literature, even though 2 models have been developed into tools to be used by parents and healthcare providers [15, 58]. In this context, digital technology may be a promising avenue for the practical application of obesity prediction models. Considering that many of the variables found to be important in published models are modifiable, they could be useful tools to inform targeted interventions. Nonetheless, it is essential that clinically relevant risk thresholds be determined for early childhood obesity prediction.

The authors have no financial or nonfinancial conflicts of interests to disclose that may be relevant to this work. The funders had no role in the study design, data collection and analysis, the decision to publish, or the preparation of this paper.