The early detection of chronic kidney disease (CKD) continues to be one of the great challenges of global public health. Traditionally, machine learning models have required large volumes of labeled data, expensive laboratory tests, and a broad set of clinical variables, limiting their deployment in community or resource-poor settings. However, an emerging approach based on large language models (LLMs) is demonstrating that it is possible to perform effective screenings without the need for specific training per dataset, using only a handful of clinical features readily obtainable in the community.
This article explores how to move from a reliance on big data to a meaningful, minimalist approach, where intelligent variable selection—guided by machine learning algorithms—allows LLMs to act as zero-shot screening tools. The results obtained with models such as LLaMA-3, Qwen-3, Mistral and GPT-4o-mini, evaluated in three heterogeneous datasets from different countries, are analyzed, demonstrating that it is possible to achieve clinically relevant performance with only a few selected variables.
The key lies in the ability of LLMs to infer patterns from textual descriptions of patients, serializing tabular records using standardized templates. This process, combined with rigorous attribute selection based on feature importance analysis, reduces complexity without sacrificing accuracy. In fact, experiments show that the reduced subset of variables not only equals, but improves balanced accuracy and probability estimates compared to the full set, offering adequate performance for population screening purposes.
For companies and organizations looking to implement AI solutions in healthcare, this approach represents a concrete opportunity. The integration of AI for enterprises through LLMs makes it possible to develop screening systems that do not rely on expensive data infrastructures or specialized labeling equipment. At Q2BSTUDIO, we understand that the true value of artificial intelligence lies in its practical and scalable applicability. That's why we offer artificial intelligence services designed to adapt to the real needs of each sector, from healthcare to industry.
One of the most innovative aspects of the study is the generalization of the model to different populations. By using a set of clinically meaningful variables available in the community (such as age, blood pressure, creatinine levels, and the presence of diabetes), LLMs demonstrate robustness to changes in distribution and population biases. This is especially relevant in environments where historical data is scarce or unrepresentative. The ability of these models to operate in zero-shot mode eliminates the need to retrain the system each time it is applied to a new region or demographic, dramatically reducing deployment and maintenance costs.
From a technical perspective, serializing tabular data into text is a critical step. The way prompt templates are structured directly influences the quality of the predictions. The researchers used standardized formats that present the patient's information to the LLM as a brief narrative description, followed by a binary question about the likelihood of suffering from CKD. This approach, although simple in appearance, requires a deep knowledge of the clinical domain and the linguistic capabilities of the model. At Q2BSTUDIO, we work with AI agents capable of processing heterogeneous information and generating contextual responses, applying similar prompting techniques optimized for classification and assisted diagnosis tasks.
ML-guided feature selection not only reduces dimensionality, but also improves model interpretability. By identifying the most relevant variables—such as the albumin-to-creatinine ratio or the estimated glomerular filtration rate—it is easier for health professionals to understand, who can validate and trust the predictions. This transparency is critical to the clinical adoption of any AI-based system. In addition, the use of accessible community features allows screening to be carried out in pharmacies, primary care centers or even through mobile applications, democratizing access to early detection.
Another relevant point is the comparison between models. Open source LLMs such as LLaMA-3 and Qwen-3 showed competitive performance against proprietary models such as GPT-4o-mini, opening the door to on-premises deployments without reliance on external cloud services. However, for environments that require scalability and high availability, combining with AWS and Azure cloud services may be the ideal solution. At Q2BSTUDIO, we integrate our AI solutions with the main cloud platforms, guaranteeing security, regulatory compliance and massive processing capacity. This is especially relevant when handling sensitive patient data, where cybersecurity must be a priority by design.
The application of this approach is not limited to CKD screening. The trait selection and zero-shot inference methodology can be extrapolated to other chronic diseases such as diabetes, hypertension or cardiovascular diseases. The versatility of LLMs allows prompt templates to be quickly adapted to new clinical contexts, provided that a reduced set of relevant variables is available. This makes LLMs a unified platform for multiple screening, reducing the fragmentation of systems that currently exists in many health systems.
For companies in the health and technology sector, this trend represents a clear business opportunity. Developing custom applications that integrate LLMs for population screening can make a difference in emerging markets or regions with limited infrastructure. At Q2BSTUDIO, we specialize in creating custom software that combines artificial intelligence with existing business processes, generating efficient and sustainable solutions. Whether it's implementing a screening system in a network of clinics or developing a telemedicine platform with predictive capabilities, our team has the necessary technical expertise.
In addition, integration with business intelligence service tools such as Power BI allows screening results to be visualised in real time, facilitating decision-making by healthcare managers. The combination of LLMs with interactive dashboards offers a complete view of the health status of a population, identifying risk areas and optimizing resource allocation. At Q2BSTUDIO, we offer power bi solutions that connect directly with predictive models, providing dynamic reporting and early warnings.
Finally, it is important to consider the ethical and privacy aspects. Using clinical data to train or infer models requires careful handling of sensitive information. The LLMs used in this study operate in zero-shot mode, meaning they do not store or retain patient data, but process each consultation independently. This reduces the risks of information leakage, although it does not eliminate them completely. Implementing robust cybersecurity measures, such as end-to-end encryption and data anonymization, is essential for any deployment in real-world environments. At Q2BSTUDIO, we design systems with advanced security protocols, complying with regulations such as GDPR and HIPAA when necessary.
In conclusion, the shift from a lot of data to a meaningful set of characteristics is redefining screening for CKD and other diseases. LLMs, combined with intelligent variable selection, offer a viable, cost-effective, and generalizable alternative to traditional machine learning methods. For organizations looking to adopt this technology, having a technology partner that understands both clinical and engineering is critical. At Q2BSTUDIO, we are prepared to accompany this process, from conceptualization to implementation, offering artificial intelligence, cloud, cybersecurity and business intelligence solutions that adapt to the specific needs of each client.
