Medical artificial intelligence could transform healthcare from a system focused on treating disease to one that prevents it in the first place, but only if researchers and clinicians close the gap between what AI can technically do and what actually improves patient outcomes. A new framework published in Frontiers in Science outlines where AI's greatest opportunities lie and how the healthcare industry can move beyond diagnosis toward earlier prediction and prevention.

Why Is AI Struggling to Move Beyond Diagnosis?

Despite rapid advances in AI technology, most clinical applications today remain stuck in advisory and decision-support roles. A 2026 Stanford-Harvard review found that nearly half of more than 500 medical AI studies relied on exam-style questions rather than real patient data, and only 5% used actual patient information. When AI systems faced uncertainty, incomplete information, or realistic clinical workflows, their performance often dropped significantly.

The core problem is straightforward: a technically accurate prediction is not automatically good medicine. An AI model can identify a pattern in seconds, but a clinician may need years of training to understand whether that pattern matters for the specific person sitting in front of them. This distinction explains both the promise and the limits of artificial intelligence in healthcare.

"Healthcare systems around the world are facing growing pressures from aging populations, rising rates of chronic disease, and workforce shortages. The need to get this right has never been greater," said Dr. Hutan Ashrafian, Senior Author and Clinical Expert at Imperial College London.

Dr. Hutan Ashrafian, Senior Author at Imperial College London

What Would a Prevention-Focused AI Healthcare System Actually Look Like?

Researchers have extended an established framework for medical AI with two new domains: Critique and Creative. This expanded model lays out a spectrum of clinical autonomy, ranging from "advisory" tools that supplement clinical judgment without being embedded in workflows, through "co-pilot" tools that share tasks with clinicians who remain in control, to "navigator" systems that work with minimal human oversight.

Almost all AI in clinical use today sits at the advisory and co-pilot end of this spectrum. However, as AI models become able to use multiple data streams simultaneously, navigator roles could emerge across personalized treatment, remote patient monitoring, clinical decision support, robotic surgery, and hospital operations. This shift could help drive a broader movement toward earlier prediction and prevention of disease, with significant implications for global population health.

One concrete example is AI-based cervical cancer screening, which has shown promise but remains inconsistently deployed across healthcare systems. Limited clinical validation, variation in clinical practice, and regulatory considerations continue to hold back widespread adoption, even though the technology works.

How to Bridge the Gap Between AI Capability and Clinical Benefit

• Prioritize Clinical Evidence Over Benchmark Scores: AI models need evaluation with real patients, clinicians, workflows, and local medical data rather than laboratory benchmarks alone. Patient outcomes, not model accuracy, should determine whether an AI application succeeds.
• Design Workflows That Support Rather Than Replace Clinician Judgment: The clinician needs to understand what an AI system recommends, why it makes that recommendation, and when not to use it. A Stanford Medicine study found that physicians working alongside a chatbot performed similarly to the chatbot alone, but workflow design mattered significantly.
• Ensure Data Quality and Local Relevance: Datasets may contain demographic gaps, coding errors, outdated information, or patterns that do not transfer to another health system. AI systems trained on one population may not work reliably for another.
• Establish Clear Lines of Responsibility: As AI becomes more deeply embedded in clinical care, clear accountability is essential, including how liability is shared between clinicians, AI developers, and hospitals.
• Address Practical Implementation Challenges: Reimbursement models, procurement pathways, institutional governance, and ongoing post-deployment monitoring are critical to real-world adoption.

Regulation alone will not be sufficient to ensure impactful adoption. Instead, addressing practical considerations and building strong understanding of how clinicians and AI systems can work together effectively will be essential.

Real-World AI in Action: Taiwan's Pharmacy AI Agent

A concrete example of AI moving into frontline healthcare is now unfolding in Taiwan. Qualcomm and ASUS have launched the "Pharmaceutical AI Agent" program, designed specifically for community pharmacists and running locally on edge devices rather than relying on cloud computing. The initiative addresses a real clinical need: Taiwan officially became a "super-aged society" in 2025, with its population aged 65 and above exceeding 4.67 million, and nearly 40% of older adults take multiple medications.

The Pharmaceutical AI Agent integrates multiple sources of medication information to help pharmacists identify potential risks such as drug-drug interactions, duplicate therapies, and inappropriate medication use. The system was optimized from a 120-billion-parameter large language model down to a 20-billion-parameter model capable of running locally on AI PCs powered by Qualcomm technology. By processing information on-device rather than sending it to the cloud, the system strengthens data privacy and security while supporting pharmacists in the demanding work of cross-checking information across multiple prescriptions.

"The potential applications of AI in medicine extend far beyond what we are currently seeing in diagnosis. AI could help us detect disease earlier, tailor treatments more precisely, and shift the focus from treating illness to preventing it in the first place. But technology alone isn't enough, and translating the promise into real-world impact has proven more challenging than many anticipated," said Dr. Ahmad Guni, Lead Author at Imperial College London.

Dr. Ahmad Guni, Lead Author at Imperial College London

The program will deploy AI PCs and edge AI boxes to more than 50 demonstration pharmacies across Chiayi, Tainan, Kaohsiung, and Pingtung, validating the feasibility, usability, and effectiveness of AI-assisted medication review in real-world settings. This approach demonstrates how sovereign AI, which keeps sensitive data local rather than transmitting it to external servers, can be put into practice in smart healthcare.

What Experts Say About the Path Forward

The authors of the Frontiers in Science article emphasize that realizing AI's full potential will require advances not only in algorithms and computing power, but also in data infrastructure, workforce readiness, clinical integration, and governance. The benefits of AI are most visible when technology addresses work limited by time, scale, or information volume. A health system may generate millions of laboratory results, medical images, clinical notes, and monitoring signals. AI algorithms can process this data and flag cases that may require attention, allowing clinicians to focus on the evidence most relevant to patient care.

"AI is attracting so much attention because it offers the possibility not only of improving clinical care, but also of reducing errors and workflow inefficiencies, easing pressures on healthcare services, and ultimately delivering much better experiences and outcomes for patients. Today's challenge is not simply developing more and more powerful AI, but ensuring that these technologies are integrated in ways that genuinely improve the quality of care for patients," said Dr. Ashrafian.

Dr. Hutan Ashrafian, Senior Author at Imperial College London

Ultimately, bridging the gap between technical capability and real-world clinical benefit demands sustained investment not only in AI innovation, but also in the governance, infrastructure, sustainability, and trust needed to support its safe and equitable adoption. The next phase of medical AI will be defined not by how powerful the algorithms are, but by how well they integrate into the actual workflows of healthcare professionals and how effectively they improve outcomes for real patients.