Abstract

Heart failure remains a pressing global health challenge, affecting approximately 64 million people and incurring annual healthcare costs of $108 billion, driven by high readmission rates and delayed interventions. Current risk prediction models predominantly rely on static machine learning approaches, utilizing fixed datasets that fail to adapt to the evolving nature of patient health. This limitation results in outdated risk assessments, hindering timely clinical responses and worsening patient outcomes and resource efficiencies. This research addresses a fundamental methodological gap by developing and validating a chunk-based ensemble framework that enables researchers to evaluate dynamic machine learning potential using readily available static clinical datasets, with synthetic heart failure data serving as a controlled proof-of-concept. Using the Heart Failure Prediction - Clinical Records Dataset from Kaggle, comprising 5,000 synthetic records with 13 clinical attributes, this study addressed two research questions on how a simulation framework can be designed using sequential patient data partitioning, and how within this controlled environment performance compares to traditional static approaches. The methodology employed a comprehensive Data Analytics approach following the CRISP-DM framework, splitting the dataset into baseline, follow-up, and testing subsets. Seven machine learning algorithms including Random Forest, XGBoost, Gradient Boosting Machine, SVM, Logistic Regression, KNN, and Naive Bayes were implemented using Python in Jupyter Notebook. The framework employed an innovative chunk-based ensemble methodology where models were trained on baseline data combined with sequential follow-up chunks, then predictions were averaged to simulate dynamic updating behaviour. The findings validate the framework's ability to detect performance differences between static and dynamic approaches within controlled conditions, though the exceptionally high-performance metrics achieved reflect the simplified nature of synthetic data rather than realistic clinical expectations. The research contributes a scalable framework for evaluating sequential patient data integration that could guide future validation studies using real-world clinical datasets. These outcomes demonstrate the methodological foundation for evaluating whether dynamic approaches could potentially enhance clinical decision-making through more responsive and accurate risk prediction, pending validation with authentic longitudinal clinical data.

Publication Date

12-9-2025

Document Type

Thesis

Student Type

Graduate

Degree Name

Professional Studies (MS)

Department, Program, or Center

Graduate Programs & Research

Advisor

Sanjay Modak

Advisor/Committee Member

Ayman Ibrahim

Comments

This thesis has been embargoed. The full-text will be available on or around 1/5/2027.

Campus

RIT Dubai

Download

Share

COinS