Project Marketplace

Project Topics & Materials

Search curated materials across every major department, faculty and institution.

Admin

About This Research Topic Modern healthcare is undergoing a profound shift as artificial intelligence takes on a growing share of diagnostic decision-making once reserved almost exclusively for clinicians. From reading medical images to flagging early markers of diabetes and cardiovascular disease, AI-powered diagnostic tools promise faster, more consistent, and more scalable disease detection. Yet impressive headline accuracy figures can conceal a more complicated reality: how a model performs on average is not the same as how it performs for every patient, every subgroup, or every clinical setting. This article unpacks a rigorous statistical evaluation of AI-based disease diagnosis systems, built around two widely used benchmark datasets. Rather than relying on a single accuracy score, the evaluation draws on a battery of statistical tools: descriptive statistics, logistic regression baselines, ROC and AUC analysis, calibration diagnostics, and subgroup fairness testing using the McNemar test. The result is a clearer picture of what AI diagnostic systems actually deliver — and where they fall short. Students and researchers working on similar quantitative or health-analytics topics may also find it useful to browse related statistics and data science project topics for inspiration on structuring their own evaluation studies. Main Abstract The rapid adoption of artificial intelligence in clinical diagnosis has outpaced the statistical scrutiny needed to confirm that these systems are safe, fair, and clinically dependable. This study carries out a wide-ranging statistical assessment of AI-based disease diagnosis systems, focusing on classification accuracy, sensitivity, specificity, the diagnostic odds ratio, area under the receiver operating characteristic curve (AUC-ROC), and calibration performance. Using secondary data drawn from two well-established public health datasets — the Pima Indians Diabetes Dataset and the UCI Heart Disease Dataset — the analysis combines descriptive statistics, hypothesis testing, correlation analysis, logistic regression, ROC curve analysis, and the McNemar test to compare model performance. Convolutional neural networks and gradient boosting classifiers outperformed a conventional logistic regression baseline, recording AUC-ROC scores of 0.91 and 0.89 respectively, against 0.76 for logistic regression. Despite this aggregate advantage, the study uncovered statistically significant disparities in sensitivity and specificity across demographic subgroups, alongside evidence of mild overconfidence in the models' probability estimates. These findings indicate that although AI diagnostic tools can outperform traditional statistical baselines on average, subgroup performance gaps and calibration weaknesses must be addressed before wide-scale clinical rollout. The study recommends fairness-aware training methods, ongoing statistical performance audits, and the incorporation of explainability techniques into clinical validation workflows.

₦5,000View

Can't find your topic? Request a custom material →