Nutrition for Precision Health, Powered by the All of Us

The Nutrition for Precision Health (NPH) study is a multi-module, large-scale project with observational and interventional components embedded in the All of Us Research Program. The overarching goal is to develop artificial intelligence and machine learning (AI/ML) algorithms that predict individual responses to different diet patterns. The study consists of three modules designed to balance breadth (large scale phenotyping) and depth (controlled dietary interventions):

Module 1: Phenotyping (non-interventional)

Module 2: Community-dwelling controlled feeding group (Intervention arm 1)

Module 3: Residential (Live-in) controlled feeding group (Intervention arm 2)

Approximately 8,000 participants are anticipated to be enrolled in Module 1. From this cohort, approximately 1,200 participants will enroll in Module 2, and a separate subset of Module 1 participants (approximately 150 participants) will enroll in Module 3. Module 1 is observational and only Modules 2 and 3 are interventional in nature (intervention arms).

Module 1 is a phenotyping observational study. Participants undergo comprehensive characterization across an 8 to 10-day baseline period for assessments including: clinical measures, biospecimen collection, wearable sensor monitoring, questionnaires, and a liquid meal test (LMT). During the LMT, participants ingest a standardized liquid meal with a dose of acetaminophen for estimating gastric emptying and provide timed blood samples for postprandial profiling. The LMT is a diagnostic stimulus used solely for feature generation and is not being evaluated as an intervention. The resulting Module 1 high dimensional dataset supports machine learning methods (e.g., PCA, clustering, recursive feature elimination) for candidate predictor discovery and is therefore not listed in the Arms/Interventions section. Module 1 data will be used to develop novel statistical and machine learning methods to learn individual and generalizable causal models of nutrition and health, particularly in the presence of missing or incomplete data. The scale of Module 1 enables discovery of causal pathways and moderators between physical and contextual measures, LMT responses, and continuous glucose monitoring (CGM) data. These insights are critical for developing models that not only predict individual dietary responses but also explain why and under what conditions they occur.

Module 2 is a community dwelling controlled feeding arm (Intervention arm 1). A subset of Module 1 participants will enroll in Module 2, in which participants consume three standardized eucaloric diets; Diet A, Diet B, Diet C, in a crossover sequence. Each diet period lasts 12-14 days, separated by a minimum of 14-day washout period between diets. All meals are provided, but participants remain in their community dwelling environments. Participants undergo one of six possible sequences of dietary interventions, reflecting all possible orderings of the three diets (ABC, ACB, BCA, BAC, CAB, and CBA). To reduce potential bias, all six diet sequences are included in the crossover design. Rather than assigning diet sequences to individual participants, a cohort-based randomization approach is used to reduce operational burden on the metabolic kitchens. In this approach, pre-generated schedules involving all 6 diet sequences are randomly assigned to the clinical site metabolic kitchens, with each diet sequence corresponding to a cohort. Participants are then enrolled into the cohorts. The six diet orders are repeated over time at each site until the full enrollment targets for Modules 2 and 3 are met. Each site follows a different randomized version of the overall cohort schedule, which ensures distribution of the possible diet orders across time and clinical sites while preserving balance and logistical feasibility. Wearable-generated data (like accelerometry and CGM), physical and contextual measures are collected throughout.

At the completion of each diet participants are provided a standardized breakfast meal test, DSMT, from each of the three provided diets (Diet A, B, C). The 0-4 hour postprandial response curves are then used to evaluate inter-individual variation and unmask features collected in Module 1 and 2 that contribute to the AI/ML predictions of the response.

Each breakfast serves as a controlled stimulus to reveal underlying biological features from rich multimodal data streams, including clinical, molecular, behavioral, environmental, and social domains that may drive interindividual variability in metabolic response. For each DSMT, participants provide blood samples for up to nine time points over four hours. Analyte concentrations of glucose, insulin, triglycerides, and GLP-1, are used to construct response curves. The primary outcomes are the area under the curve (AUC) for each analyte (glucose, insulin, triglycerides, GLP-1) following each DSMT. This approach parallels a cardiac stress test: the stimulus (the meal) is a probe to expose individual variability in physiological function. Module 2 data will also be used to develop novel statistical and machine learning methods that produce individual and generalizable causal models of nutrition and health.

Module 3 is a controlled feeding study (Intervention arm 2) that is implemented in the live-in/residential setting. A separate subset of Module 1 participants completes the same three diets (Diet A, B, C) as in Module 2, but while residing in a research setting with full supervision of intake, activity, and sleep. The same cohort randomization scheme as in Module 2 determines diet order. The residential environment in Module 3 provides the highest degree of experimental control, allowing isolation of physiological effects of diet composition. Participants undergo a diet-specific meal test (DSMT) as well as a liquid meal test (LMT) accompanied by a dose of acetaminophen after completing each of the diets. In addition, intake balance studies are conducted in this module using the doubly labeled water assessments and DXA for body composition. Data from Module 3 will help quantify and separate variance in the AI/ML models attributable to adherence and other community dwelling factors observed in Module 2 and enables rigorous and controlled comparison against causal relationships discovered in Module 2 data.

In both Modules 2 and 3, participants are masked to the nutritional profile of each diet to minimize expectancy bias. Investigators and diet implementation staff are unmasked.