ASN Scientific Sessions

Predictive Modeling in Nutrition

How does predictive modeling influence nutrition research, clinical interventions, and public health? This question was pondered in a session co-chaired by David Allison, PhD, and Kevin Hall, PhD, on Saturday morning at Experimental Biology. Dr. Allison introduced the session by giving an overview of some uses of predictive modeling related to nutrition research: modeling the social contagion of obesity, obesity propagation over generations, projecting policy implications, and agent-based modeling.

Kevin Hall, PhD, shared his influential work that started with developing predictive models and now uses them within clinical studies. Using previous highly controlled studies in metabolic chambers, he was able to form a mathematical model to predict metabolic adaptations in response to changes in energy intake. Recently, he and colleagues have used these models to design and predict what occurs in response to reducing the level of carbohydrate in the diet. The carbohydrate-insulin hypothesis, which postulates that carbohydrates cause body fat gain through the increase of insulin, has recently gained prominence from those filling in research gaps with their own speculations. To address this gap, he tested the predictions of his mathematical model against this hypothesis in a metabolic ward study by providing isocaloric diets restricted in either carbohydrates or fat. Though the low carbohydrate diet group did exhibit reduced insulin, both groups lost body fat, and in fact, the reduced fat group had a greater fat loss,because energy expenditure was decreased by the reduced carb group. In response to this study, Dr. Hall noted that they were criticized by some low carb advocates that their study wasn’t “true low-carb”, and that they didn’t wait long enough to observe a rebound in energy expenditure on a low-carb diet. However, in 2016 they published a second study on a ketogenic diet. The small, but not clinically significant increase in energy expenditure on the ketogenic diet was predicted by the model. Such successful application of mathematical modeling demonstrates that the models can be used to make testable predictions for experiments, and data can be continuously integrated to evolve the model.

Corby Martin, PhD, talked about the application of predictive modeling into the clinical setting. Combining such modeling into technology can theoretically improve adherence to a weight loss intervention, in part by allowing feedback between the clinician and the patient. Dr. Martin emphasized however that the presence of a device does little itself to change behavior over the long-term; it must be used in combination with a behavior change theory and the goals of the patient. He and his colleagues have been developing mobile health interventions to bring the rigor a clinic-based weight loss intervention to larger scale. For example, their SmartLoss app allows tracking of objective data such as weight that can be be fed into a predictive model to quantify diet adherence, and feedback can be sent to the patient in near real-time. Such data can be useful to individualize recommendations; for instance patterns may arise if patients don’t come often enough for clinic visits, or if certain methods of calorie restriction like portion control work better than others. Further, it can make it easy to visually show the plateau effect of even small intermittent non-adherence. Their initial trial demonstrated a 9.4% weight loss using the app versus 0.6% in those only receiving health tips by smartphone.

Next, Ben van Ommen, PhD, discussed the concept of “personalized nutrition” – that health is defined by “the ability to adapt”, and that we all have variability within certain oxidative, inflammatory, and metabolic processes in response to dietary challenges. In other words, we have a phenotypic flexibility in the ability to adapt to inflammation, or different amounts of nutrients in the diet (e.g. fats vs. carbs). This may be indicated by Blanco-Rojo et al 2015, who found that a low-fat versus Mediterranean diet may result in improved insulin sensitivity depending on whether participants have liver or muscle insulin resistance. Dr. van Ommen also discussed the issue of data ownership, something that they’ve put a lot of thought into in Europe that will become a bigger issue as we move toward personalization of nutrition.

Lastly, Emily Dhurandhar, PhD, talked about the usefulness of predictive modeling for public health obesity policies. Two issues with public health policies is that they often don’t account for energy balance being a dynamic, adaptable system, because the models are built in clinical settings, and they often focus on one food or habit instead of achieving a substantial shift in overall energy balance. For instance, Dr. Hall and colleagues showed that using the simple 3500 kcal per pound fat gain rule greatly overestimates how much weight will be lost over time in response to calorie reductions because it fails to account for metabolic adaptations in energy expenditure. In addition, behavioral compensation is not considered in health policies. Dhurandhar and colleagues showed that such compensation in exercise interventions to increase energy expenditure may be 55-64% less than expected because of such compensation like increased appetite. From this data they built an “E-EBALANCE” calculator that corrects for behavioral compensation. Dr. Hall’s group also revealed the underappreciated role of appetite in slowing weight loss compared to energy expenditure. If models that inform public health policies don’t take both metabolic and behavioral adaptations into account, the effectiveness of public health recommendations will be overestimated.

Diana Thomas summarized the session by concluding that predictive modeling can be used to ask “what if” questions, and these results can form the basis for targeted experimental design, which lends more rigor to the experiment. It can be used to determine deviations from expectation during weight loss interventions, and closely assess patient adherence, while guiding behavior change. We are slowly uncovering evidence of individual metabolic responses to dietary challenges, and predictive modeling will have a role in the future of personalized nutrition. Modeling needs to be incorporated into assessing the potential impacts of public policy, and tracking unintended changes and long-term national trends. It is clear that modeling will play an important role in nutrition research, the clinic, and public policy going forward.