Body mass index multiple regression formula testable by eight Bradford Hill causality criteria: Worldwide ecological cohort data analysis to inform dietary guidance

BackgroundArtificial intelligence (AI) analytics have not been applied to global burden of disease (GBD) risk factor data to study population health. The comparative risk assessment (CRA) systematic literature review-based methodology for population attributable fractions (PAFs in percents) calculations has not been utilised for quantifying dietary and other risk factors for body mass index kg/M2 (BMI). MethodsInstitute of Health Metrics and Evaluation (IHME) staff and volunteer collaborators analysed over 12,000 GBD risk factor surveys of people from 195 countries and synthesized the data into representative mean cohort BMI and risk factor values. We formatted IHME GBD data relevant to BMI and associated risk factors. We empirically explored the univariate and multiple regression correlations of BMI risk factors with worldwide BMI to derive a BMI multiple regression formula (BMI formula). Main outcome measures included the performances of the BMI formula when tested with all nine Bradford Hill causality criteria each scored on a 0-5 scale: 0=negative to 5=very strong support. FindingsThe BMI formula derived, with all foods in kilocalories/day (kcal/day), BMI formula risk factor coefficients were adjusted to equate with their PAFs. BMI increasing foods had "+" signs and BMI decreasing foods "-" signs. Total BMI formula PAF=80.96%. BMI formula=(0.37%*processed meat + 4.23%*red meat + 0.02%*fish + 2.24%*milk + 5.67%*poultry + 1.77%*eggs + 0.34%*alcohol + 0.99%*sugary beverages + 0.04%*corn + 0.72%*potatoes + 8.48%*saturated fatty acids + 3.89%*polyunsaturated fatty acids + 0.27%*trans fatty acids - 2.99%*fruit - 4.07%*vegetables - 0.37%*nuts and seeds - 0.45%*whole grains - 1.49%*legumes - 8.62%*rice - 0.10%*sweet potatoes - 7.45% physical activity (METs/week) - 20.38%*child underweight + 6.02%*sex (male=1, female=2))*0.05012 + 21.77. BMI formula versus BMI: r=0.907, 95% CI: 0.903 to 0.911, p<0.0001. Bradford Hill causality criteria test scores (0-5): (1) strength=5, (2) experimentation=5, (3) consistency=5, (4) dose-response=5, (5) temporality=5, (6) analogy=4, (7), plausibility=5, (8) specificity=5, and (9) coherence=5. Total score=44/45. InterpretationNine Bradford Hill causality criteria strongly supported a causal relationship between the BMI formula derived and mean BMIs of worldwide cohorts. The artificial intelligence methodology introduced could inform individual, clinical, and public health strategies regarding overweight/obesity prevention/treatment and other health outcomes. FundingNone Research in contextO_ST_ABSEvidence before this studyC_ST_ABSComparative risk assessment (CRA) systematic literature review-based methodology has been used in worldwide global burden of disease (GBD) analysis to determine population attributable fraction(s) (PAF(s)) for one or more risk factors for various health outcomes. So far, CRA has not been applied to derive PAFs for dietary and other risk factors for worldwide BMI. Artificial intelligence (AI) analytics has not yet been applied to worldwide GBD data as an alternative to the CRA methodology for determining risk factor PAFs for health outcomes. Added value of this study{square}A multiple regression derived BMI formula (BMI formula) including PAFs of 20 dietary risk factors, physical activity, childhood severe underweight, and sex satisfied all nine Bradford Hill causality criteria. The BMI formula also plausibly predicted the long-term BMI outcomes related to various dietary and physical activity scenarios. All the BMI formulas 24 risk factor PAFs were consistent in sign (+ or -) with the preponderance of previously published studies on those risk factors related to BMI. Implications of all the available evidenceThe AI analytics methodology of GBD data modeling of BMI and associated risk factors infers causality of the BMI formula estimates with BMI worldwide and BMIs of subsets. This methodology may enable multiple regression formulas for risk factors of health outcomes for a range of non-communicable diseases--testable by Bradford Hill causality criteria.