Is a score enough? Pitfalls and solutions for AI severity scores

Area of Science:

• Radiology
• Artificial Intelligence (AI)
• Psychological Science
• Statistics

Background:

• Artificial intelligence (AI) tools in radiology commonly provide severity scores, indicating pathology likelihood.
• The utility and transparency of these AI-generated scores remain under-examined.
• Existing research has not sufficiently addressed the radiologist-AI interaction.

Purpose of the Study:

• To elucidate six human factors limitations of AI scores in radiology.
• To propose a hypothesis for mitigating these limitations.
• To discuss empirical testing of the proposed hypothesis.

Main Methods:

• Analysis of AI score utility drawing on principles from psychological science and statistics.
• Identification of six key human factors limitations: inter-AI variability, intra-AI variability, inter-radiologist variability, intra-radiologist variability, unknown score distribution, and perceptual challenges.
• Formulation of a hypothesis involving false discovery rate (FDR) and false omission rate (FOR).

Main Results:

• Six human factors limitations were identified that undermine the utility of AI severity scores.
• These limitations include variability across and within AI systems and radiologists, unknown score distributions, and perceptual challenges.
• A hypothesis is proposed: FDR and FOR thresholds can mitigate these limitations.

Conclusions:

• The utility of AI severity scores in radiology is significantly constrained by human factors.
• Addressing variability and perceptual challenges is crucial for effective radiologist-AI interaction.
• Incorporating FDR and FOR as thresholds offers a potential strategy to enhance the reliability and utility of AI scores.