Innovations in early detection of chronic non-communicable diseases among adolescents through an easy-to-Use AutoML paradigm

In this research, we present an interpretable AutoML approach for the early diagnosis of hypertension and hyperinsulinemia among adolescents, conditions that are critical to identify during these formative years due to their requirement for lifelong care and monitoring. The dataset, collected from 2...

Full description

Saved in:
Bibliographic Details
Published in:Health care management science Vol. 28; no. 3; pp. 434 - 460
Main Authors: Rankovic, Nevena, Rankovic, Dragica, Lukic, Igor
Format: Journal Article
Language:English
Published: New York Springer US 01.09.2025
Springer Nature B.V
Subjects:
Accuracy
Adolescence
Adolescent
Business and Management
Child development
Chronic Disease
Chronic illnesses
Cross-Sectional Studies
Deep learning
Diabetes
Disease
Early Diagnosis
Econometrics
Epidemics
Female
Health Administration
Health care industry
Health Informatics
Humans
Hypertension
Hypertension - diagnosis
Insulin resistance
Lifestyles
Machine learning
Male
Management
Metabolic syndrome
Neural networks
Noncommunicable Diseases - diagnosis
Operations Research/Decision Theory
Patients
Regression analysis
Serbia
Support vector machines
Teenagers
Serbia
AutoML paradigm
Tabular variational autoencoder (TVAE) synthetic generation
Model-agnostic/post-hoc methods
Chronic non-communicable diseases
ISSN:1386-9620, 1572-9389, 1572-9389
Online Access:Get full text
Tags: Add Tag
No Tags, Be the first to tag this record!
Description
Summary:In this research, we present an interpretable AutoML approach for the early diagnosis of hypertension and hyperinsulinemia among adolescents, conditions that are critical to identify during these formative years due to their requirement for lifelong care and monitoring. The dataset, collected from 2019 to 2022 by Serbia’s Healthcare Center through an observational cross-sectional study, posed challenges common to medical datasets, including imbalances, data scarcity, and a need for transparent, explainable predictive models. To counter these issues, we utilized three AutoML frameworks - AutoGluon, H2O, and MLJAR - in conjunction with a Tabular Variational Autoencoder (TVAE) to synthetically augment the data points, Prinicipal Component Analysis (PCA) for dimensionality reduction, and SHapley Additive exPlanations (SHAP) and Permutation feature importance analyses to extract insights from the results. AutoGluon outperformed the others on the original dataset, delivering better results with weighted ensemble models for both conditions under a 12-minute budget-time constraint and maintaining all evaluation metrics below a 4% threshold, all without the need for further scaling or calibration in the experimental setup. Our research underscores the broad applicability of the current AutoML paradigm, highlighting its particular benefits for the healthcare domain and diagnostics, where such advanced tools can enhance patient care.
Bibliography:ObjectType-Article-1
SourceType-Scholarly Journals-1
ObjectType-Feature-2
content type line 14
content type line 23
ISSN:1386-9620
1572-9389
1572-9389
DOI:10.1007/s10729-025-09718-6