Developing and validating machine learning algorithms to predict various indices of diet quality among a socioeconomically disadvantaged group

Public health research faces challenges in recruiting hard-to-reach populations, including socioeconomically disadvantaged groups. This study evaluated whether machine learning (ML) algorithms developed using data from a general population could predict indices of diet quality among a socioeconomically disadvantaged group. Data from 5367 adults (77.5% females) in the NutriQuébec project and on 122 variables potentially associated with dietary intakes were used. Dietary intakes were measured using a validated web-based 24-h recall. Participants were categorized by fifths of a deprivation score based on income, education, and material and social deprivation. Participants in the first 4 fifths were assigned to the general NutriQuébec sample (n= 4180) and those above the 5th quintile formed the high deprivation sample (n= 1187). Three indices of diet quality defined as "high" or "low" were used: vegetable and fruit consumption (VFC, ≥5.0 reference amounts [RA]/day), "other foods" consumption, i.e. foods not recommended in Canada's Food Guide 2019 (OFC, >5.0 RA/day), and overall diet quality measured using the Healthy Eating Food Index-2019 (HEFI-2019, >48.9 points). The algorithms developed and tested in the general NutriQuébec sample predicted high VFC, OFC and HEFI-2019 with accuracies of 0.60 (95%CI:0.58-0.62), 0.58 (95%CI:0.56-0.60) and 0.61 (95%CI:0.59-0.63), respectively. In the high deprivation sample, the algorithms predicted the diet quality indices with comparable accuracies (VFC, 0.69, 95%CI:0.67-0.71; OFC, 0.56, 95%CI:0.54-0.58; HEFI-2019, 0.66, 95%CI:0.65-0.67). ML algorithms trained to predict three diet quality indices in the general NutriQuébec sample were applicable to a sample with high deprivation. Additional studies should further evaluate applicability to other underrepresented groups.