Artificial intelligence-driven population health management improving healthcare value & equity
Health systems globally are faced with failed ethical commitment to their patients and financial extinction if they fail to consistently provide clinically efficacious, societally equitable, cost-effective healthcare. Despite the known causal link between the nutrition-related chronic disease epidemics and the world’s top morbidity cause, cardiovascular disease, there is no evidence-based, cost effective, scalable model of nutrition education intervention for and with medical trainees and professionals and their patients. Similarly, there is no known demonstrated case successfully applying artificial intelligence (AI)driven Big Data within a population health management framework for such an intervention to optimally refine it. Therefore, the medical school-based teaching kitchen, The Goldring Center for Culinary Medicine (GCCM) at Tulane University School of Medicine, launched the largest known multi-site cohort study with nested Bayesian adaptive randomized controlled trial (BA-RCTs) across 30 medical centers and 3,785 medical trainees/professionals. Cooking for Health Optimization with Patients (CHOP) with its four sub-studies features not only the first known systematic review and metaanalysis on this subject to determine best practices. CHOP also serves as the first known nutrition education study utilizing the latest AI-based machine learning (ML) techniques to complement the traditional statistical approaches to provide real-time, precise treatment estimates for causal inference and assessment of hands-on cooking and nutrition education for medical professionals and trainees’ patient counseling competencies, and improved patient psychometric and biometric outcomes. (1) The first sub-study, CHOP-Meta-analysis, demonstrated that though the average effect size (ES) across the 10 eligible nutrition education studies among medical trainees was 10.36 (95%CI 6.87-13.85; p<0.001), the only study meeting the STROBE criteria for high quality, the phase I sub-study of CHOP-Medical Students below, had significantly triple the ES (31.67; 95%CI 29.91-33.43). (2) CHOP-Medical Students demonstrated in inverse variance-weighted fixed effects meta-analysis of propensity score-adjusted fixed effects multivariable regression across 2,982 students that GCCM versus traditional clinical education significantly improved trainees’ total mastery counseling patients in 25 nutrition topics (OR 1.64; 95%CI 1.53-1.76; p<0.001). (3) CHOP-CME demonstrated that among 230 medical professionals, GCCM education significantly increased these odds, but by 159% more than the trainees’ improvement (OR 2.66; 95%CI 2.26-3.14; p<0.001) in addition to significantly increasing the odds of counseling most patients on nutrition in their clinical practices (OR 5.56; 95%CI 2.124-14.18; p<0.001). (4) CHOPCommunity demonstrated that GCCM education versus the standard of care significantly increased patient adherence to the Mediterranean diet (MedDiet) (OR 1.94; 1.04-3.60; p=0.038) and greater connectedness in their social networks (p=0.007). The pilot RCT for diabetes patients, CHOP-Diabetes, nested in this sub-study demonstrated superior improvements in diastolic blood pressure (-4 versus 7 mmHg, p=0.037) and cholesterol (14 versus 17 mg/dL, p=0.044) for patients randomized to GCCM versus the standard of care. The nested Phase II BA-RCT, CHOP-Family, demonstrated that GCCM versus standard of care had significantly greater MedDiet adherence based on their grocery receipts (OR 4.92; 95%CI 1.78-13.56; p=0.002). Using the Random Forest Multiple Imputation ML algorithm, the simulated Phase III BA-RCT predicted 93 hospital admissions and $3.9 million would be saved providing GCCM versus standard of care for congestive heart failure (CHF) exacerbation-risk patients primarily from underserved communities. Among 41 tested ML algorithms, the top performing Iterative Classifier Optimizer was comparable to the estimated traditional statistical model for the trainees’ primary endpoint for (1) (RMSE 0.314 versus 0.282), and the top performing Kstar was superior to the traditional model for the professionals’ primary endpoint (RMSE 0.431 versus 0.414). The four sub-studies within CHOP taken together provide the first known multi-site cohort and BA-RCT evidence for superiority of hands-on cooking and nutrition education compared to the standard of education and medical care for improved trainee/professional nutrition counseling competencies and patient outcomes. CHOP utilized the state-of-the-art in causal inference-based statistics, randomized trials for causal assessment, and ML to provide robust, precise estimates of comparative treatment effectiveness. This research infrastructure was scaled up to meet GCCM’s growing programmatic needs as it has since grown over 5 years to 30+ medical centers providing 53,674+ teaching hours to 4,171+ medical trainees/professionals and patients. CHOP has utilized the latest rigorous study design and analysis methodologies to provide a blueprint to optimize health systems through sustainable improvements as population health management that is clinically and cost effective, reducing health inequities while improving individual outcomes.