Biases and Trustworthiness Challenges with Mitigation Strategies for Large Language Models in Healthcare

Rapid innovations in Large Language Models (LLMs) have resulted in remarkably efficient decision-making and learning capacities, specifically in critical sectors such as healthcare. Domain-specific LLMs are progressively being designed for medical pre-screening and diagnostic procedures in healthcare. Despite these advancements, LLMs persist as opaque systems lacking the capacity to offer fair decisions and trustworthy explanations. Though various techniques are proposed to address challenges associated with LLMs, further research is considered essential to adopt LLMs in high-risk sectors. This article explores challenges related to LLMs along with suggested strategies for mitigation. Among several challenges, this study presents a comprehensive overview of biases and trustworthiness in healthcare LLMs. It presents an overview of clinical, cognitive, and demographic bias mitigation approaches at the Data, Model, and Inference levels. Further, it provides a detailed critical analysis of existing bias quantification metrics and healthcare benchmarks to assess trustworthiness in clinical LLMs. This research is supported by an empirical study, where existing patient records are extracted to fine-tune Llama2. The responses from fine-tuned Llama 2 are analyzed for various biases, and existing bias mitigation strategies are applied. However, the results indicate existing bias mitigation approaches need to be revised, highlighting the need for advanced techniques. This study concludes by explaining the essential research areas in bias mitigation and trustworthiness necessary to ensure the practical deployment of LLMs in clinical decision-making.