Operationalizing fairness in medical AI adoption:Detection of early Alzheimer’s Disease with 2D CNN
Heising L.M., & Angelopoulos, S.
Objectives: To operationalize fairness in the adoption of medical artificial intelligence (AI) algorithms in terms of access to computational resources, the proposed approach is based on a two-dimensional (2D) Convolutional Neural Networks (CNN), which provides a faster, cheaper, and accurate-enough detection of early Alzheimer’s Disease (AD) and Mild Cognitive Impairment (MCI), without the need for use of large training datasets or costly high-performance computing (HPC) infrastructures.
Methods: The standardized ADNI datasets are used for the proposed model, with additional skull stripping, using the BET2 approach. The 2D CNN architecture is based on LeNet-5, the LReLU activation function and a Sigmoid function were used, and batch normalization was added after every convolutional layer to stabilize the learning process. The model was optimized by manually tuning all its hyperparameters.
Results: The model was evaluated in terms of accuracy, recall, precision, and f1-score. The results demonstrate that the model predicted MCI with an accuracy of .735, passing the random guessing baseline of .521, and predicted AD with an accuracy of .837, passing the random guessing baseline of .536.
Discussion: The proposed approach can assist clinicians in the early diagnosis of AD and MCI, with high-enough accuracy, based on relatively smaller datasets, and without the need of HPC infrastructures. Such an approach can alleviate disparities and operationalize fairness in the adoption of medical algorithms.
Conclusion: Medical AI algorithms should not be focused solely on accuracy but should also be evaluated with respect to how they might impact disparities and operationalize fairness in their adoption.
Suggested citation: Heising, L.M., and Angelopoulos, S. (2022). Operationalizing fairness in medical AI adoption: Detection of early Alzheimer’s Disease with 2D CNN, BMJ Health & Care Informatics, 29:e100485. DOI: 10.1136/bmjhci-2021-100485.