In May 2021, a county-level hospitals CT room in China welcomed a “new colleague” - an AI-assisted diagnostic system. This tireless assistant could instantly identify suspicious lesions in CT images, providing doctors with a “second opinion” on diagnosis. However, subsequent data revealed a puzzling phenomenon: doctors‘ diagnostic reports became more detailed, but the number of CT scans the department could process daily decreased significantly.

This discovery comes from a research team led by YANG Zheshuai, a researcher under the “Hundred Talents Program” at the School of Management, Zhejiang University. Through in-depth analysis of thoracic CT diagnostic data from the hospital, they revealed a paradox in the field of digital healthcare: while AI assistance improves the quality of doctors‘ work, it may reduce work efficiency. This research shifts the focus from patient perception to clinician behavior and is of key significance for optimizing the allocation and management of medical resources, ultimately improving patient satisfaction and well-being. It also offers important insights for grassroots hospitals that are actively promoting digital transformation.

In recent years, the application of artificial intelligence (AI) in the medical field has experienced explosive growth, with its clinical applications now covering multiple medical areas. From IBM Watson‘s oncology system to Google DeepMind‘s diagnosis of retinal diseases, AI technology has demonstrated its enormous potential in improving diagnostic accuracy and efficiency. By 2025, it is projected that 90% of hospitals will be using medical AI to assist doctors in their work.

However, most studies focus on large tertiary hospitals or top-tier physician groups, paying little attention to the effectiveness of AI in grassroots hospitals with relatively limited resources and generally average physician qualifications. This is precisely why YANG Zheshuai‘s team chose county-level central hospitals as their research subjects. The hospitals selected in this study area serve as the main medical centers of the county, receiving many patients daily, and the overall qualifications of their CT department physicians are at an average level, representing the typical situation of many grassroots hospitals in China.

“Grassroots hospitals are the capillaries of China‘s healthcare system, serving the largest patient population,” the study points out. “Understanding the real impact of AI in these environments is crucial for optimizing the allocation of healthcare resources nationwide.”

The research team collected thoracic CT diagnostic data from the hospital 180 days before and after the system went live, analyzing the impact of AI introduction on doctors‘ work performance. The results revealed a seemingly contradictory phenomenon:

■   Significantly improved work quality

With the introduction of AI assistance, the length of CT reports written by doctors has significantly increased - both the conclusion and descriptive sections have grown considerably, resulting in a more detailed and comprehensive diagnostic process. This is because AI‘s superior image processing accuracy can identify sub-visual abnormalities and small lesions that are often missed by the human eye, thereby reducing diagnostic omissions and improving the comprehensiveness of detection. Moreover, AI‘s data-driven learning can provide expert-level “second opinions,” which is particularly beneficial for junior clinicians. In addition, AI can promote adherence to operational procedures, reducing non-standard practices and subjective errors.

■   Work efficiency has declined significantly

Contrary to expectations, AI assistance did not improve doctors‘ work efficiency. Data shows that the total number of chest CT reports processed daily in the CT department has decreased significantly, with an average reduction in the number of reports processed per doctor per day. This finding challenges the common assumption that “AI can improve efficiency.”

This “quality-efficiency” trade-off reveals the complex impact of AI assistance in real-world medical scenarios. More importantly, this relationship is not static but evolves dynamically over time. The research team further analyzed monthly data within six months of AI introduction and discovered a clear trend. Specifically, improvements in work quality steadily increased, while declines in work efficiency gradually intensified. In other words, as the use of AI assistance extends, the problem of declining efficiency worsens over time.

Why did AI assistance fail to achieve the expected “win-win” situation? The research team pointed out that three key factors led to the decrease in efficiency.

First, AI increases the cognitive burden on junior radiologists. Less experienced physicians may experience self-doubt when their diagnoses differ from the AI‘s output, leading them to spend more time verifying and confirming, thus prolonging the diagnostic time for individual cases. Second, a limited understanding of AI systems reduces their trust in them. Doctors‘ trust in AI systems is not built overnight; without a full understanding of the AI‘s decision-making logic, they must spend more time verifying results to ensure patient safety. Third, AI assistance alters existing clinical workflows, requiring doctors to explore optimal modes of collaboration with AI. During this adaptation period, a temporary decrease in efficiency is almost inevitable.

It is worth noting that these mechanisms may be more pronounced among physicians of average qualifications. They lack the extensive experience of senior specialists to quickly validate AI suggestions, yet they may also develop higher self-expectations due to AI assistance, leading them to prioritize quality over efficiency. This reflects the real dilemma faced by primary healthcare institutions - they aspire to improve service levels through new technologies but are constrained by both resources and capabilities.

Faced with the “quality-efficiency” paradox brought about by AI assistance, how should hospital administrators and AI developers respond? Based on their findings, the research team has put forward several practical suggestions.

■   Differentiated application strategy

Hospitals should develop differentiated AI usage strategies based on the complexity and risk level of cases. For complex, high-risk cases, AI assistance should be prioritized to ensure diagnostic quality; for routine cases, AI use can be appropriately limited to maintain overall work efficiency.

■   Strengthen doctor training

Training content should not be limited to AI system operation, but should also include how to quickly interpret AI suggestions, integrate them into the diagnostic process, and maintain independent clinical judgment with AI assistance. This will help doctors build reasonable trust in the AI system and reduce unnecessary verification time.

■   Optimize system design

AI developers should focus on creating more user-friendly interfaces, providing transparency and explainability in decision-making, and helping doctors understand the basis of AI‘s judgments. Continuous updating and expanding the database, optimizing algorithms, and improving diagnostic accuracy will enhance doctors‘ trust in the system.

This study provides valuable insights for hospitals at the grassroots level. In general, when introducing AI systems, potential efficiency reductions should be anticipated, and corresponding strategies should be developed. At the same time, recognizing that doctors need a considerable amount of time to adapt to AI-assisted work, administrators should provide sufficient support and training to help teams transition smoothly.

Research has shattered our simplistic notions of AI-assisted healthcare scenarios, revealing the complexity of applying digital technologies in real-world medical settings. AI is not a panacea; its effectiveness is influenced by multiple factors, including organizational environment, user characteristics, and implementation strategies.
In resource-constrained primary healthcare settings, the “quality-efficiency” trade-off presented by AI assistance deserves particular attention. It reminds us that technological innovation must be accompanied by organizational change and personnel training to truly realize its value. In the future, as AI technology matures and doctors‘ adaptability improves, this paradox may gradually ease. However, at the current stage, recognizing this challenge and actively seeking a balance is key to promoting the development of digital healthcare.

- We thank Researcher YANG Zheshuai and the research team for their valuable contribution to advancing the understanding of how AI-assisted diagnostics influence physicians’ work quality and efficiency in grassroots healthcare settings, offering important insights into the quality-efficiency trade-off and the real-world challenges of digital transformation in medical practice.

- You can read the original article in Chinese  here