Unmasking AI bias in traditional prognosis models

Abstract

Artificial Intelligence (AI) in healthcare is on the rise, assisting with early disease prognosis and decision-making processes. However, A? models exhibit a propensity for bias, which can lead to unequal performance across various demographic and regional subgroups. Recent advancements in metalearning aim to enhance the generalisation of these models when faced with limited data scenarios. Yet, there is a paucity of studies investigating the behaviour of such models within low-resource clinical datasets, particularly when factoring in fairness constraints. This research employs sophisticated metalearning methodologies to assess and quantify bias in predicting rabies outcomes by implementing ProtoNet, Model-Agnostic Meta-Learning (MAML), and a Hybrid ProtoMAML model to a custom-annotated rabies dataset that has been stratified by age, gender, and district. The results demonstrate that the ProtoMAML model achieves superior fairness; however, bias remains evident across subgroups. This study underscores the importance of addressing algorithmic bias within clinical AI applications and advocates for developing equitable and inclusive medical decision-making systems.