The STRIPED Dietary Supplement Label Explorer: A Tool to Identify Supplements Sold with Weight-Loss, Muscle-Building, and Cleanse/Detox Claims

Abstract

Background: Limited federal premarket oversight over United States-sold dietary supplements impedes consumer safety and product efficacy. The Dietary Supplement Label Database (DSLD) was created to increase publicly available information on United States-sold dietary supplements. Building on what the DSLD was designed to provide, we aimed to create a comprehensive database that can facilitate searches on supplements sold with weight loss, muscle building, and cleanse/detox claims, supplement categories previously flagged for misleading claims and containing toxic ingredients.

Objectives: This study aims to leverage publicly available DSLD Application Programming Interface (API) to develop an easy-to-use tool to classify DSLD supplement labels with weight loss, muscle building and cleanse/detox claims.

Methods: A 4-step categorization methodology was used to develop the tool: 1) create reference standard database by deductively coding claims (weight loss, muscle building, and cleanse/detox) on 5000 DSLD labels; 2) develop 3 systematic heuristics (1 per claim) and refine heuristics as assessed by recall, specificity, precision, negative predictive value, F1 Score, and accuracy; 3) develop multimodal deep learning model as an additional method to identify the 3 claims; and 4) compare models' performance using the receiver operating characteristic (ROC) curve and efficiency analyses (i.e. hours of human labor taken to develop each model).

Results: Of the 4745 DSLD labels included in the reference standard database, 4.2% were defined using the criteria as weight loss, 6.3% muscle building, and 3.0% cleanse/detox. Three systematic heuristics for each claim were refined 4 times, with pass 4 exceeding prior passes' performances. ROC curve analyses indicated that systematic heuristic performed significantly better (P < 0.05) than the multimodal deep learning model at classifying cleanse/detox labels, yet efficiency analyses found systematic heuristics less efficient (110 compared with 30 h).

Conclusions: Our findings illustrate the feasibility of using the DSLD API to create a tool that classifies weight loss, muscle building, and cleanse/detox labels using our supplement label categorization methodology. This publicly available tool, STRIPED Dietary Supplement Label Explorer, may be used to support future research and the monitoring of claims on dietary supplement labels.

Keywords: Dietary Supplement Label Database; dietary supplements; label database; nutraceuticals; supplement labels.