HealthGuard - A Comparative Supervised Learning System for Disease Risk Prediction Across Multiple Healthcare Datasets

Pullen, Ben and Secco, Emanuele Lindo HealthGuard - A Comparative Supervised Learning System for Disease Risk Prediction Across Multiple Healthcare Datasets. Medinformatics. ISSN 3029-1321 (Accepted for Publication)

Text 10301_v10.pdf - Accepted Version Restricted to Repository staff only until 1 March 2027. Download (1MB)

Abstract

This paper presents HealthGuard, an Android health application that delivers two side-by-side artificial intelligence (AI) risk classifiers - one for coronary heart disease and one for type 2 diabetes - behind a shared safety layer. Most consumer mobile health applications either embed a single classifier or confine themselves to manual logging; this work demonstrates an architectural pattern for hosting multiple disease-specific classifiers within a single application without blurring their clinical contexts. Two classification pipelines were deployed, and within each pipeline four candidate algorithms - Logistic Regression, Decision Tree, Random Forest, and Gradient Boosting - were compared during model selection. The heart pipeline was trained on the UCI Cleveland Heart Disease dataset and the diabetes pipeline on the Pima Indians Diabetes dataset, both publicly available benchmark sets. Model selection used GridSearchCV with stratified 5-fold cross-validation, and the winner of each pipeline was re-evaluated on a held-out 20% test set. Logistic Regression was selected for both pipelines: on Cleveland it achieved a cross-validated Receiver Operating Characteristic Area Under the Curve (ROC-AUC) of 0.9070 +/- 0.0114 and a test ROC-AUC of 0.9589, and on Pima a cross-validated ROC-AUC of 0.8439 +/- 0.0156 and a test ROC-AUC of 0.8098. The distinctive contributions of this work are a condition-aware safety layer that attaches plain-language explanations grounded in published clinical thresholds, together with a comparative evaluation of how surfaces - rather than hides - the performance gap between the two pipelines and the disjoint feature-importance rankings their selected models produce.

Item Type:	Article
Faculty / Department:	Faculty of Human and Digital Sciences > School of Computer Science and the Environment
Depositing User:	Emanuele Secco
Date Deposited:	30 Oct 2026 10:16
Last Modified:	30 Oct 2026 10:16
URI:	https://hira.hope.ac.uk/id/eprint/5006

Actions (login required)

View Item