Beyond accuracy: a framework for evaluating algorithmic bias and performance, applied to automated sleep scoring.

Area of Science:

• Computational neuroscience
• Medical artificial intelligence
• Sleep medicine

Background:

• Artificial intelligence (AI) has advanced sleep-scoring algorithms to near-theoretical performance limits.
• Clinical adoption is hindered by validation, fairness, and oversight requirements.
• Current validation methods like Bland-Altman analysis may miss biases influenced by external factors.

Purpose of the Study:

• To propose a universal framework for quantifying performance metrics and biases in AI predictive tools.
• To extend validation methods to analyze the full distribution of performance and errors influenced by external factors.
• To apply this framework to U-Sleep and YASA algorithms for sleep apnea risk assessment.

Main Methods:

• Developed a framework analyzing how external factors influence the entire distribution of predictive performance and errors.
• Extended conventional validation beyond mean-based metrics to include quantiles and bias analysis.
• Applied the framework to U-Sleep and YASA algorithms using sleep-scoring data.

Main Results:

• Identified biases in U-Sleep and YASA, such as age-related shifts, suggesting data imbalances or missing information.
• Demonstrated that both algorithms maintain non-inferior performance in sleep apnea risk assessment despite identified biases.
• The proposed framework quantifies biases and performance across the entire distribution, offering deeper insights than mean-based metrics.

Conclusions:

• The universal framework provides a more comprehensive validation of AI predictive tools, addressing clinical adoption challenges.
• AI sleep-scoring algorithms show clinical utility in risk assessment, even with identified biases.
• Further research should focus on addressing identified biases to enhance AI algorithm reliability and fairness in clinical settings.