Wrist Photoplethysmography Predicts Dietary Information

TL;DR

The study investigates whether wrist photoplethysmography (PPG) captures dietary information by aligning PPG-derived representations with meal descriptions using a multimodal Nutrition PPG Language Model (NPLM). The approach leverages near-meal PPG within a four-hour window and a linear adapter to map PPG embeddings to GPT-2 token space, enabling downstream predictions of daily caloric intake and postprandial satiety. Across a real-world (AHMS) and a controlled dining (Validation Study) cohort, PPG signals near meals improved predictive performance by about 11% in AUC for intake and enhanced satiety prediction, with robustness to degraded text and generalization to an independent dataset. These findings suggest wearable PPG can support passive dietary monitoring and motivate future work on metabolic endpoints and integration with clinical workflows.

Abstract

Whether wearable photoplethysmography (PPG) contains dietary information remains unknown. We trained a language model on 1.1M meals to predict meal descriptions from PPG, aligning PPG to text. PPG nontrivially predicts meal content; predictability decreases for PPGs farther from meals. This transfers to dietary tasks: PPG increases AUC by 11% for intake and satiety across held-out and independent cohorts, with gains robust to text degradation. Wearable PPG may enable passive dietary monitoring.

Figures (4)

• Figure 1: Alignment of PPG with language reveals a nutritional signal.(a) Diagram of the NPLM architecture. Modality-specific encoders project inputs into a shared language embedding space. Alignment is trained by maximizing meal-text likelihood conditioned on near-meal PPG. (b) Evidence of Alignment. Left: Changes in win rate when conditioning on time-matched randomly permuted PPG segments within-subject versus randomly permuted across-subject. Right: Changes in win rate for PPG segments trained at increasing time lags relative to the logged meal within subject; the win rate monotonically decreases.
• Figure 2: NPLM predicts dietary outcomes across multiple experimental conditions.(a)Predicting dietary outcomes across cohorts. We report the area under the receiver operating characteristic curve (AUC) for daily caloric intake classification (AHMS) and postprandial satiety prediction (Validation Study), comparing multimodal NPLM against text-only baselines. (b)Relationship between performance and text length in AHMS. Removing text lowers average model performance. Text alone outperforms PPG alone, however, PPG with 50% of the tokens outperforms full text. When removing all text (100%), the results illustrate that PPG alone has limited information, though it is powerful when added to (simplified) meal description. (c)Performance with semantic summarizations of food logs. NPLM with summarized meal descriptions outperforms the original full-text baseline. (d)Surrogate nutrient model coefficients.
• Figure 3: Screenshots of the Validation Study app. The app allows participants to log their appetite and access study information.
• Figure 4: Histogram of meal diversity scores across all individuals in the AHMS cohort. The meal diversity score is defined as the number of unique meals eaten divided by the total number of meals in a given month. The average meal diversity score across all individuals is 0.54. Each participant is weighted equally in this distribution, regardless of the total number of meals logged (inclusion criterion: at least 5 days with at least one meal logged per day).