A Review on Human-Centered Machine Learning for Decision Support in High-Stakes Domains

Abstract

Machine learning is increasingly used to support important decisions in healthcare, public services, finance, and aviation. These systems can help with prediction, risk assessment, resource allocation, and professional decision-making. However, high accuracy does not always mean that a system is safe, fair, easy to understand, or suitable for real use. This paper reviews human-centered machine learning for decision support in high - stakes domains. It discusses common machine learning methods and several impo rtant human-centered requirements, including explainability, appropriate trust, usability, workflow integration, fairness, accountability, and human oversight. It also compares these requirements across different fields. Current studies show that many syst ems perform well in experiments or on historical data, but there is still limited evidence from real-world use. Explanations may increase trust, but they do not always improve decision quality. In some cases, they may even cause people to rely too much on wrong AI advice. Different fields also have different priorities. Healthcare focuses on patient safety, public services focus on fairness, finance focuses on rules and appeals, while aviation and manufacturing focus more on reliability, situation awareness, and skill retention. This paper argues that machine learning should support human judgment instead of replacing it.