Digitized medical services scale up, but real-world performance remains uneven

As artificial intelligence gains traction in China's healthcare system, experts say its expansion is following a clear pattern, shaped as much by clinical realities as by technological capability.

Across academia and industry, there is a growing consensus that AI healthcare holds tangible potential in improving efficiency, extending access, and shifting care toward earlier intervention. Yet where and how these systems are deployed remains highly selective.

"AI tends to be adopted first in areas where data is standardized, workloads are heavy, and risks are relatively controllable," said Wu You, an associate professor at Tsinghua University's medical school.

"These conditions make it easier to integrate AI as a tool for assistance and quality control," Wu said.

That logic has placed imaging-related specialties at the forefront. In radiology, AI systems are increasingly used to assist with CT and MRI interpretation-tasks that are repetitive, high-frequency, and governed by standardized workflows.

Digital pathology is another fast-growing area, supporting the development of regional pathology centers and cloud-based diagnostic networks, although human experts still dominate most clinical decisions.

Beyond diagnostics, AI is also gaining ground in hospital quality management. Tools for prescription review, medical record auditing, and disease-specific quality control are being deployed to improve consistency and reduce error rates.

Meanwhile, patient-facing applications, such as intelligent triage, pre-consultation services, and automated billing, are reshaping hospital workflows, even if many rely more on digital infrastructure than advanced clinical intelligence.

Despite this progress, adoption remains concentrated in top-tier hospitals.

"In most cases, AI systems are first piloted in large tertiary hospitals, which are deeply involved in their development and validation," Wu said. "Only after that do they gradually expand to regional healthcare networks. Direct adoption at the primary-care level is still limited."

One structural reason is that the current medical practice still assumes that physicians must be capable of evaluating and, when necessary, correcting AI-generated outputs. In lower-tier settings, where technical training and resources may be more constrained, that condition is not always met.

Another challenge lies in the gap between controlled testing and real-world performance.

Regulatory approval in China typically confirms that AI systems meet baseline standards for safety and accuracy under controlled conditions. But once deployed in clinical environments, outcomes can vary widely.

"Effectiveness depends on who is using the system, how it is integrated into workflows, and how long it has been in use," Wu said.

In a multiyear study tracking AI-assisted lung CT analysis at two hospitals in Beijing, Wu and her colleagues observed markedly different patterns between junior and senior physicians.

Less experienced doctors were more likely to rely on AI suggestions, sometimes leading to short-term increases in false positives and longer reporting times. Senior doctors, by contrast, showed more stable performance, with gradual improvements emerging over time as workflows adapted.

The findings suggest that early declines in efficiency, particularly among less experienced users, should not necessarily be interpreted as technological failure, but rather as part of the adjustment process in human-machine collaboration.

More broadly, they point to the limits of a regulatory approach focused solely on market entry.

"In high-risk and complex medical scenarios, governance cannot stop at approval," Wu said. "It needs to extend to continuous evaluation and adaptation in real-world use."

As AI becomes more deeply embedded in healthcare systems, experts say the next phase will depend less on technological breakthroughs than on how effectively these systems are integrated, monitored, and governed.

The promise of AI in medicine may be substantial. But its success, they argue, will ultimately be determined outside of controlled trials, in the messy, variable realities of everyday clinical practice.