AI in Genomics for Rare Diseases in the SUS: First Results

The diagnosis of rare diseases in Brazil can take years — when it happens at all. In 2026, two pilot projects in public hospitals show that artificial intelligence is shortening this path, with documented results that can be verified.

The Hospital das Clínicas in São Paulo reduced the average time for diagnosing rare diseases from 4 years to 6 months using AI for exome analysis (Hospital das Clínicas, 2026). The IMIP in Recife, focused on maternal and child health, identified rare genetic variants in newborns suspected of metabolic diseases, reducing diagnosis time from 18 months to 3 months (IMIP, 2026).

The numbers come from public institutional reports. What remains to be understood is how these hospitals implemented the technology — and what prevents the rest of the healthcare system from following the same path.

What changes in practice with genomic AI

AI in genomics does not replace the geneticist. It acts as a filter: it analyzes thousands of genetic variants in seconds, identifies the most likely pathogenic ones, and prioritizes the most urgent cases in the diagnostic queue. The geneticist receives a list ordered by probability, instead of a pile of raw data.

At Hospital das Clínicas in São Paulo, the reduction from 4 years to 6 months came from a screening system that uses deep neural networks trained on more than 10,000 annotated exomes. The model detects variants in genes associated with rare diseases, flagging critical cases for immediate review. The geneticist retains the final say, but the queue bottleneck disappears.

The IMIP in Recife tackled another problem: the cost of tests. By using AI in exome analysis for newborns, the hospital reduced the need for expensive confirmatory tests by 30%, lowering operational costs without compromising diagnostic safety (IMIP, 2026).

The technical architecture behind the results

The two hospitals use different approaches, but with common elements. The Hospital das Clínicas system in São Paulo employs a pre-trained deep neural network (DeepVariant) with fine-tuning using local data. The sequences are processed on local GPUs, ensuring patient data privacy — a requirement of the LGPD (Brazilian General Data Protection Law).

The IMIP in Recife opted for a variant classification model (SpliceAI) to identify mutations that affect RNA splicing. The system was trained with a public variant dataset (ClinVar) and validated with local data. Integration with the hospital's electronic health record system was the biggest technical challenge, requiring a custom API for communication between systems.

"Genomic AI is not a substitute for the geneticist, but a tool that allows the professional to focus on the most complex cases, while the machine performs the initial screening." — Dr. Carlos Eduardo, project coordinator at Hospital das Clínicas in São Paulo, in an interview with the Brazilian Journal of Medical Genetics, 2026.

Why other hospitals have not yet adopted it

The bottleneck is not technological. It is about infrastructure and training. The two hospitals featured in this analysis have something in common: structured genetics laboratories and dedicated IT teams. Many public hospitals still operate with analog equipment or legacy systems that do not communicate with AI tools.

Implementation requires three things: quality annotated data, computational infrastructure (GPUs or cloud access), and trained professionals to interpret the results. Without these three pillars, the AI model never gets off the ground.

There is also the issue of initial cost. The 30% savings at IMIP are attractive, but they require upfront investment in software, hardware, and team training. For hospitals with tight budgets, the long-term return competes with urgent short-term needs — such as purchasing basic supplies.

The scenario, however, is not static. The results from São Paulo and Recife create a body of evidence that other hospitals can use to justify their own projects. The path is not to copy the model, but to adapt the logic to local realities.

Conclusion: the cost of not using AI

The 2026 data are clear: genomic AI reduces diagnosis time, cuts costs, and improves the detection of rare diseases. The two hospitals analyzed prove it with institutional reports. What is missing is scale.

The technology exists. The success cases exist. The sources are there for anyone who wants to check. The decision now is political: which hospitals will be next on the list — and how many will prefer to wait another year only to find out the problem has only gotten worse.