Outcome-based approaches

Two recent studies have explored PN effects in terms of increased weight loss from following a so-called New Nordic Diet instead of following a standard Western diet for half a year based on either a single phenotypic (fasting plasma glucose) or a single genetic biomarker (the ratio between Prevotella and Bacteroides gut bacteria)^{(Reference Ritz, Astrup and Larsen12,Reference Hjorth, Christensen and Larsen13)} . Specifically, the increased weight loss obtained from following the New Nordic diet compared with following the standard Western diet for 26 weeks was shown, using a linear mixed model with a biomarker-diet interaction, to depend linearly on the baseline fasting plasma glucose level: the higher the glucose level at the baseline, the larger the additional weight loss due to the New Nordic diet after 26 weeks^{(Reference Ritz, Astrup and Larsen12)}.

In principle, the outcome-based approach might utilise many different types of data simultaneously and not just a single biomarker. However, if the approach relies on structure imposed through regression modelling as was the case in the two above examples, then using more data becomes much more challenging and indeed a topic of ongoing research in the statistical literature^{(Reference Shen and Cai11)}. Moving beyond regression models towards grey- and black box modelling using other data analytic tools such as machine learning and random forests would be an interesting alternative. Sticking to fairly simple regression models also has its charm as it would allow easy comparison between studies or even enable combining results across studies and it would also help ensure reproducibility, two aspects that may be useful at present in advancing PN.

Outcome-based approaches rely heavily on empirical or mechanistic models that have already established the link between biomarker information and the outcome of interest. Data from RCTs, which have already been conducted, may be revisited and re-analysed (in various ways) to provide such models. Thus, in a first step, it does not require running new trials as already existing data can be exploited, potentially providing novel insights at a low cost. Moreover, whether or not the RCT resulted in a positive result in the first place is not important, as there could be biomarker-specific effects in any case^{(Reference Matsouaka, Li and Cai10)}. RCTs could also be planned to investigate PN effects directly, but to our knowledge, no such RCT has yet been conducted.

In analogy to personalised medicine, outcome-based approaches could also be extended to provide binary decision rules for when to recommend one diet over another diet for achieving a certain improvement.

Population reference approaches

Recently, two studies applied population reference approaches for predicting glycaemic response and plasma glucose, respectively, based on individual data including food intake^{(Reference Berry, Valdes and Drew14,Reference Mendes-Soares, Raveh-Sadka and Azulay15)} . Specifically, individual glycaemic and postprandial responses were predicted from individual data, such as baseline characteristics, single-nucleotide polymorphisms, gut-microbiome features and habitual diet, by means of multivariate linear regression and a machine learning method (boosting), respectively. Predicted responses were correlated against observed responses, but no decision rules for new altered diets were derived.

However, in several other recent studies, decision rules were applied: Individuals would receive expert- and/or machine learning-based personal counselling as compared with standard ‘one fits all’ advice in order to change dietary habits^{(Reference Celis-Morales, Livingstone and Marsaux16–Reference Bianchi, Mariotti and Lluch18)}. For instance, if fat intake was (predicted to be) higher than recommended, the personalised counselling suggested changes in the diet aiming to reduce fat intake, such as ‘Limit your intake of saturated fats found in butter, full-fat dairy products and processed foods e.g. biscuits, pastries and processed meats’^{(Reference Celis-Morales, Livingstone and Marsaux16)}.

There are very many options for deriving data-driven decision rules as the ones used in some of the above-mentioned articles. In principle, it may be any method capable of classification, including statistical regression models, random forests and machine learning methods. There is no easy way subsequently to unravel the separate effects (on any specific outcome) resulting from these pieces of PN dietary advice. For each individual, it would require very detailed and meticulous bookkeeping of all pieces of advice provided during the time period in question, including time stamps, duration and applied decision rule(s). Subsequently, it would also require a diligent use of statistical methods to quantify the effects of pieces of personalised dietary advice. The more individual data were used for providing the PN dietary advice, the more difficult it will be to unravel the separate effects. Consequently, repeating an RCT involving a population reference approach will not necessarily provide comparable results as different individuals will most likely receive different sequences of PN dietary advice.

Perspective

In PN, population reference approaches seem to be preferred over outcome-based approaches. One reason may be that population reference approaches seem capable of utilising vast amounts of data but are presently not so principled, and the generalisation of results is difficult. There is, however, a need to better characterise and understand the PN effects that are presently reported, as a way of improving reproducibility and generalizability. On the other hand, outcome-based approaches utilise much less data but are more principled and provide results that seem to generalise more readily. This simplicity comes at the price of restrictive modelling assumptions and the limited use of individual data. Therefore, there is a need for more flexible extensions.

Finally, presently, the implementations of both types of approaches ignore that PN is about benefits or effects at an individual level, whereas applied data analytic methods were developed for estimating and evaluating population-level effects. Consequently, these methods cannot be used alone if results should be applicable at an individual level rather than a population level. Instead, they need to be augmented or modified to become useful. This means that the classical statistical inference paradigm must be replaced by a much more recent statistical learning framework, which, among other things, involves the use of concepts such as prediction models and cross-validation^{(Reference Hastie, Tibshirani and Friedman19,Reference Varadhan, Segal and Boyd20)} .