Evaluating performance bias in face-to-BMI vision transformer models across diverse human populations
Computer vision models that estimate body mass index (BMI) from facial features offer a non-invasive, low-cost alternative to physical measurement, with uses in telemedicine, emergency care where a scale or measuring tools arent available, automated self-monitoring, and large-scale epidemiological research. Most of these models, however, are trained on government records, social media images, and…
Body mass index (BMI) estimation models derived from facial images present a non-invasive, affordable alternative for various medical and research applications. However, these models frequently suffer from dataset biases due to their training on limited sources like government records, social media, and celebrity photos, which do not adequately represent the global population.
This investigation assesses the generalization capabilities of a face-to-BMI machine learning model across different populations, focusing on the impact of morphological diversity and population-specific training data on cross-cultural accuracy. Four Indigenous groups from Malaysia, Southern Africa, the Philippines, and Bolivia were utilized to train and assess Vision Transformer (ViT-H/14) models.
The study compared four training methodologies, ranging from single-population models to models trained on an extensive global dataset. In-distribution training consistently yielded the best results, with models trained on the morphology of the target population accurately predicting BMI for that specific group. When the target population differed from the training sample, incorporating additional cross-cultural variation in the training dataset enhanced predictions for out-of-distribution populations.
Consequently, it is preferable to train models on data exclusive to the target population when available. Nevertheless, training on data encompassing a broad spectrum of human morphologies is the most effective alternative when such data is not accessible. The research concludes that while utilizing target population training data yields the most accurate outcomes, incorporating training on datasets that encompass global morphological variation substantially enhances performance in underrepresented groups.
Hence, diverse training data is crucial for creating machine learning health tools that reliably generalize across diverse human populations.
Written by urgent.news from bioRxiv's reporting — not their text. Machine-written — may contain errors; check the original before relying on it.
