Generative Artificial Intelligence in Healthcare: Applications and Limitations
Generative Artificial Intelligence in Healthcare: Practical Applications and Scientific Evidence
A Generative Artificial Intelligence in Healthcare It can organize clinical records, support case analysis, and propose molecular structures for research. To understand its contribution, it is worth examining how these applications work and what results the studies present.
Research published in journals such as Nature Medicine e JAMA Network Open They are investigating the development of natural language processing models and their use by physicians. Although the studies point to significant potential, the benefits vary depending on the task and how the models are used.
Therefore, understanding this technology requires knowledge of its applications, the available evidence, and the necessary precautions in each context.
What is generative artificial intelligence?
A generative artificial intelligence It brings together computational models that learn patterns and relationships in large volumes of data. Based on what they learn, these models generate text, images, and other representations according to the information and instructions they receive.
In healthcare, for example, these models can generate summaries of medical records and propose molecular structures for drug development. However, each result depends on the training data, the context provided, and the tool’s intended purpose.
The large language models, or LLMs (Large Language Models), work with language. To construct a response, they calculate probable sequences of text units, called tokens. As for the multimodal models They allow you to combine different formats, such as text and images.
The difference becomes clearer with an example: an AI tool can analyze clinical data and calculate the risk of a disease. Generative AI, on the other hand, can transform the information in a medical record into a summary for the medical team.
In this summary, however, the system may omit an allergy or include information that does not appear in the original document. Therefore, the healthcare professional must review the content before using it. The WHO explains these applications and their risks in the Guidelines on Multimodal Models in Healthcare.
How can generative artificial intelligence in healthcare help?
Applications include documentation, clinical support, and pharmaceutical research. However, each activity requires its own evaluation criteria.
To summarize a medical record, for example, the tool must preserve the information in the record. In drug research, on the other hand, scientists need to determine whether they can produce and test the proposed molecules.
Organization of Information and Documentation
Reviewing a patient's medical history may require reviewing appointments, test results, and prescriptions scattered across multiple documents. In this context, generative AI can compile the key data into a summary and make it easier for the team to review.
In addition, the tool can adapt the language of technical content. A text describing a specific procedure, for example, can serve as the basis for an explanation that is more accessible to the patient. The responsible professional then reviews the version for accuracy and clarity.
A World Health Organization (WHO) It lists documentation of consultations, administrative tasks, and professional development among the possible applications of these models.
Support for Clinical Reasoning
In addition to organizing documents, language models can help professionals explore hypotheses and analyze alternative courses of action. However, researchers are still investigating in which situations this support improves performance.
One study published in the Nature Medicine, in February 2025, evaluated 92 doctors in simulated clinical situations. Participants with access to GPT-4 and conventional resources had an average score of 6.5 percentage points higher in tasks involving reasoning about behavior than those who used only conventional resources.
Despite the higher scores, the group also took longer to complete the cases. Since the evaluation took place in a simulated setting, the authors emphasized the need to investigate the results in real-world clinical practice.
Generative Artificial Intelligence in Drug Development
In pharmaceutical research, generative models can propose molecular structures with characteristics of interest. Scientists then select which compounds warrant further investigation and assess the feasibility of synthesizing them in the laboratory.
After synthesis, the experiments assess properties such as biological activity and toxicity. In this way, the team tests whether the computational proposals have the appropriate characteristics to move forward with development.
One example is the rentosertib, a drug candidate for idiopathic pulmonary fibrosis, a disease that causes scarring of the lung tissue and makes it difficult to breathe.
One Phase 2a trial published in the Nature Medicine, in June 2025, evaluated the compound developed with the help of generative AI. The study included 71 participants over 12 weeks and compared different doses with a placebo.
Thus, this case illustrates the transition from a computational proposal to clinical research. The technology contributed to the discovery and design of the candidate, while the experimental studies made it possible to evaluate its behavior and effects.
What are the limitations of generative artificial intelligence in healthcare?
Scientific results vary depending on the activity. For example, a essay published in the JAMA Network Open, in October 2024, evaluated 50 doctors who analyzed clinical cases with or without access to a language model.
In this study, access to the tool did not statistically significantly improve diagnostic reasoning compared with conventional methods. Therefore, making the technology available does not, in and of itself, guarantee a benefit.
The two studies involving physicians presented in this article evaluated different tasks. Therefore, their results raise an important question: For which activity, with which users, and under what conditions did the tool prove useful?
Incorrect Answers and Biases
An answer may seem convincing but contain incorrect information. For this reason, professionals need to verify the data, consult the original documents, and check the cited references.
In addition, models may replicate imbalances in the training data. If that data underrepresents certain patient groups, for example, the tool may perform differently across populations.
A The WHO warns of these risks and of overreliance on automation. Therefore, the review must assess both the accuracy of the response and its appropriateness to the situation.
Information Protection
Medical records and research documents may contain personal data, medical histories, and confidential information. Therefore, the institution must be familiar with the tool’s terms of use before submitting these materials.
This analysis should clarify where the system stores the data, who can access it, and whether the vendor will be able to use it to train models. With this information, the organization can determine which documents it can use and what controls it needs to implement.
In addition to choosing the right technology, teams need clear guidelines for their daily routines. This way, the procedures help prevent information from being sent incorrectly.
What should you consider before adopting a tool?
In January 2026, the FDA and EMA announced Ten Joint Principles of AI Best Practices in Drug Development. The initiative provides guidance on developing practices to generate evidence using these technologies, but does not constitute an endorsement of specific tools.
In practice, the evaluation should start with a specific need. If the team intends to summarize documents, for example, it can compare the generated texts with the original records and measure errors, omissions, and the time required for review.
Before implementation, it’s worth asking:
- What problem is the tool designed to solve?
- Have the available studies evaluated a similar activity?
- What errors did the tests identify?
- Who will review the results?
- How will the team monitor performance over time?
This allows the institution to assess whether the solution contributes to its work and what adjustments it needs to make before expanding its use.
How can we turn possibilities into benefits?
A Generative Artificial Intelligence in Healthcare It expands the ways in which knowledge can be organized and scientific proposals explored. However, its contribution depends on a combination of computational power, experimental evaluation, and the participation of specialized professionals.
In the pharmaceutical industry, a proposal gains value when the team is able to investigate it thoroughly. To do so, it must bring together the protocols, documentation, people, and materials needed for its implementation. Learn more How the Supply of Medications for Clinical Research Works.
At FIC Company, we support this continuity with logistics solutions for the pharmaceutical industry and clinical research, taking into account the requirements of each project. If your operation involves the supply, storage, or transportation of medications and supplies, Talk to our team to identify needs and plan the next steps. Solutions that advance science.
Share on:
Se inscrever
Login
0 Comments
Mais Votado
Mais Recente
Mais Antigo
Feedbacks Inline
Ver todos os comentários