AACE 2026: Machine learning model predicts insulin resistance

A machine learning algorithm using routinely collected nonfasting clinical variables predicted insulin resistance with strong performance across independent cohorts and projected substantial 5-year health care cost savings according to research abstract presented at the 2026 AACE Annual Meeting in Las Vegas.

Using National Health and Nutrition Examination Survey (NHANES) 2017 to 2020 data from 4,625 adults with complete homeostatic model assessment for insulin resistance (HOMA-IR) measurements, investigators trained three supervised machine learning models—logistic regression, random forest, and gradient boosting—to classify insulin resistance, defined as HOMA-IR greater than 2.5.  

Nonfasting Predictors

The models incorporated 11 nonfasting predictors:

1. age,

2. sex,

3. race and ethnicity,

4. body mass index (BMI),

5. waist circumference,

6. triglycerides,

7. total cholesterol,

8. high-density lipoprotein cholesterol,

9. alanine aminotransferase,

10. aspartate aminotransferase, and

11. gamma-glutamyl transferase. 

Gradient boosting achieved the strongest internal validation performance with an area under the receiver operating characteristic curve (AUROC) of 0.84, comparable to random forest at 0.84 and higher than logistic regression at 0.82. Temporal external validation using NHANES 2013 to 2016 data from approximately 4,400 adults demonstrated maintained performance with an AUROC of 0.82, supporting generalizability across independent cohorts. 

Feature importance analysis showed that BMI accounted for 44% of the predictive signal, followed by triglycerides at 20%, waist circumference at 16%, and age at 10%. Together, those variables explained more than 89% of the model’s predictive contribution. Mean predicted insulin resistance probability across the cohort was 0.53. 

Investigators also modeled the projected economic burden associated with insulin resistance progression. Estimated 5-year per-patient costs were $14,751 without intervention vs $10,072 with modeled risk reduction, yielding projected savings of $4,679 per patient. Among 1,226 patients categorized as high risk, defined as insulin resistance probability of at least 0.80, projected savings increased to $7,980 per patient. 

The researchers estimated that implementing a hypothetical screening program for 1,000 adults could generate approximately $4.68 million in 5-year cost avoidance. 

The researchers noted that assessing insulin resistance typically requires fasting glucose and insulin measurements to calculate HOMA-IR, limiting broader screening opportunities in primary care and endocrinology settings. By contrast, the proposed model relied exclusively on routinely available nonfasting variables. 

The researchers wrote that the nonfasting machine learning algorithm’s “simplicity, generalizability, and substantial projected economic benefits highlight its potential for opportunistic metabolic screening in primary care and endocrinology,” while noting that prospective validation and real-world cost-effectiveness analyses remain necessary. 

The research was led by investigators from TriHealth, Integris Health, and independent public health researchers. Funding sources and conflict-of-interest disclosures were not reported. 

(Editor's Note: These findings are from a conference presentation on an abstract and should be considered preliminary.)