Paradox of Medical IA Implementation: Why haven't advanced models transformed medicine yet?

Artificial intelligence (AI) has already demonstrated the ability to detect breast cancer in early stages and even predict disease years in advance, but the promise of a widespread revolution in medicine has not yet come true. In the article "The Paradox of Medical AI Implementation", published in its Substack Ground Truths, renowned cardiologist and researcher Eric Topol distinguishes the paradoxical contrast between the success of deep learning models (DL) in specific examinations — such as FDA-approved mammograms — and the slow adoption of modern-day Generative Large Language Models (LLMs), which promise to transform from diagnosis to hospital administration.

The FDA approved medical IA: what is there already?

The U.S. regulatory agency has already approved hundreds of AI-based devices, especially in radiology. Topol cites studies in which LD models surpass human radiologists in the interpretation of mammograms, reducing false positives and increasing the rate of early detection. For example, a trained algorithm with thousands of images was able to identify malignant nodules with 94% accuracy, against 88% of experts. However, despite the approvals, the integration of these tools into the clinical routine is still fragmented. Only a fraction of American hospitals have adopted large-scale AI systems — and the scenario is not very different in Brazil, where barriers such as cost, infrastructure and team training slow down implementation.

Medical AI Paradox: Why don't generative LLMs take off?

If deep learning already shows results in specific tasks, LLMs — such as GPT-4 and its competitors — expand the horizon for broader functions: writing reports, screening patients, supporting clinical decisions and even conversations with patients. The paradox pointed out by Topol is that, although LLMs have the potential to solve problems that DL does not achieve, they face even greater obstacles. The lack of transparency in training data, the risks of hallucination (false but convincing responses) and the difficulty of validation in real clinical environments make its adoption slower. Meanwhile, simpler and more focused DL models remain the safe bet for imaging-based diagnoses, as recently showed by a study of IA detects signs of pancreatic cancer up to three years before diagnosis.

Major obstacle to implementing medical AI in practice

The main bottleneck, according to Topol, is not technical, but cultural and operational. The trust of doctors, the integration with legacy systems of electronic medical records, regulation and legal responsibility (who answers if AI err?) create an ecosystem of inertia. In addition, the absence of clear evaluation standards for LLMs in the health area makes hospitals and clinics hesitate to adopt them. While the FDA has a clearer way to approve LD devices — comparable to traditional tests — generative LLMs require new safety and efficacy metrics that are still being defined. Topol's original article can be read in full in Ground Truths.

Frequently asked questions about medical AI

Can the AI replace radiologists yet?

Nope. AI is a diagnostic support tool, not a substitute. Studies show that the combination of AI + human radiologist achieves the best results.

Are LLMs safe for clinical use?

There's still no consensus. Models like GPT-4 can make serious errors (hallucinations) and their validation in real clinical scenarios is insufficient for large-scale approval.

When are we going to see AI generalized in medicine?

Adoption will be gradual and uneven. It is expected that, in the next 5 to 10 years, AI will be consolidated in areas such as radiology and pathology, but for broader functions (such as screening and documentation) the deadline may be longer.