Diabetes is a global health crisis, affecting over half a billion adults worldwide. More alarmingly, nearly half of cases globally, and one in four cases in the US, remain undiagnosed. Untreated diabetes can lead to severe complications such as heart disease, kidney failure, and vision loss, underscoring the need for easily accessible screening methods. Early detection is challenging because diabetes often presents with no symptoms in its initial stages, and many individuals avoid invasive blood tests. This study proposes a non-invasive, questionnaire-based approach to Type 2 Diabetes detection using Machine Learning models. Models were trained on the Diabetes Health Indicators Dataset, derived from a CDC survey, which included questions related to BMI, age, and self-reported health ratings. This method provides a scalable, low-cost screening tool. Three models were evaluated — Logistic Regression, Random Forest, and Neural Network. To minimize false negatives (i.e. missed diabetic cases) due to their severe health consequences, two novel metrics were introduced, Cost-Weighted Accuracy and Cost-Weighted Error Rate, which incorporated a cost ratio to reflect the higher ‘cost’ – developing severe health complications - of false negatives, as opposed to the relatively lower cost of false positives. Results showed that the Neural Network model achieved the highest sensitivity (86.66%) and the highest cost-weighted accuracy (75.88%), outperforming the Random Forest and Logistic Regression models.
Anthony Lau (Wed,) studied this question.