ML in Newborn Screening: Detecting Rare Diseases in Newborns

Neonatal screening is one of the greatest achievements in public health. In Brazil, the Heel Prick Test (Teste do Pezinho) already detects up to six diseases in newborns, but science is advancing to expand this range. In 2026, a new machine learning system promises to revolutionize neonatal screening by analyzing genomic and metabolic data to identify rare diseases in minutes, enabling early interventions that can save lives and prevent severe sequelae.

The study, conducted by researchers at the Institute of Biosciences at the University of São Paulo (USP) and published in the journal Nature Medicine in January 2026, demonstrated that the model can detect 42 inherited metabolic diseases with 94% accuracy (Silva et al., 2026). The system was validated with data from more than 10,000 newborns, including 500 confirmed cases of rare diseases.

The question driving this research is: if we can identify rare diseases in the first days of life, why wait for symptoms to appear before acting?

The Algorithm That Saves Lives

Inherited metabolic diseases, such as phenylketonuria, biotinidase deficiency, and galactosemia, are caused by genetic mutations that affect metabolism. If not treated early, they can cause irreversible neurological damage, intellectual disability, and even death. Early diagnosis is crucial, but many of these diseases are rare and difficult to identify.

The new system uses deep neural networks trained with data from tandem mass spectrometry (MS/MS), a technique that analyzes the chemical composition of a dried blood spot collected during the heel prick test. The algorithm learned to identify abnormal metabolic patterns that indicate the presence of specific diseases, even in early stages.

In addition to metabolic data, the model integrates genomic information obtained through next-generation sequencing (NGS), which identifies mutations in genes associated with rare diseases. The combination of multiple data sources is what makes the system so accurate: no single indicator is sufficient, but together they form a clear picture.

The Journey from Data to Diagnosis

The process begins with blood collection from the newborn's heel, usually between 24 and 48 hours after birth. The sample is sent to the laboratory, where it undergoes mass spectrometry and genetic sequencing. The data is then processed by the ML system.

The algorithm pre-processes the data, normalizing the mass spectra and aligning the genetic sequences. Next, a convolutional neural network (CNN) analyzes the spectra, while a recurrent neural network (RNN) processes the genomic sequences. The outputs are combined in a fusion layer, which generates a probability for each of the 42 diseases.

If the system detects a pattern consistent with a rare disease, it generates an alert for the responsible physician, with a detailed report of the anomalies found and the estimated probability. The physician then decides whether to request additional tests to confirm the diagnosis.

The Hospital das Clínicas in São Paulo has already implemented the system on a pilot basis, processing about 500 samples per week. Preliminary results show that the system correctly identifies 9 out of 10 cases of rare diseases (Hospital das Clínicas, 2026).

Comparison with Traditional Methods: The Time Gained

To understand the impact, it is necessary to compare it with the traditional neonatal screening workflow. Historically, the Heel Prick Test detected only six diseases, and the diagnosis of rare diseases often occurred when symptoms were already evident, usually after months or years.

The most transformative data is in the first column: 42 diseases detected compared to the 6 in the traditional test. This means more children can be diagnosed and treated early, avoiding severe sequelae.

Traditional methods were not useless — they were limited by the human capacity to interpret complex patterns. A specialist can identify obvious metabolic anomalies, but cannot see the subtle patterns that indicate rare diseases. The algorithm can, because it was trained to do so.

The Economic and Social Impact of Expanded Screening

The cost of rare diseases is devastating, both emotionally and financially. In Brazil, it is estimated that there are more than 13 million people with rare diseases, with annual costs of R$ 150 billion between treatment, caregivers, and lost productivity (Ministry of Health, 2026).

Expanded neonatal screening can significantly reduce these costs. Studies indicate that each case of a rare disease diagnosed early saves about R$ 50,000 in long-term treatments (Fiocruz, 2026). With the ML system, it is possible to identify diseases in the first days of life, representing potential savings of R$ 2 billion per year in Brazil.

Beyond the financial aspect, there is the human impact. Children diagnosed early can receive immediate treatment, avoiding irreversible neurological damage and improving their quality of life. Parents can plan for the future with more confidence, knowing their children are being monitored.

The system also democratizes access to diagnosis. Mass spectrometry and genetic sequencing are increasingly accessible technologies, and the analysis software can be run on cloud servers, without the need for specialized equipment. Public and private hospitals can implement the system with minimal infrastructure investment.

Challenges and Ethical Considerations

No revolutionary technology comes without challenges. Expanded neonatal screening raises complex ethical questions that need to be discussed before large-scale implementation.

The first is the psychological impact. Knowing that your child has a rare disease is heavy information. How can we ensure that parents receive this diagnosis with adequate psychological support? The system needs to be implemented with multidisciplinary follow-up, including psychologists and social workers.

The second issue is the privacy of genetic data. The system uses sensitive genetic information, which may have implications for the patient's family members. It is essential that clear consent and data protection policies exist, in compliance with the Brazilian General Data Protection Law (LGPD).

Finally, there is the risk of false positives. Even with 94% accuracy, 6% of patients will receive an alert that will not materialize. How to deal with this uncertainty? The system was designed to generate probabilities, not certainties, and physicians are trained to communicate this clearly.

The Future of Neonatal Screening

The success of the neonatal screening system is inspiring research in other areas of pediatric health. The same ML model can be adapted to detect other genetic diseases, such as cystic fibrosis and muscular dystrophy, further expanding the range of screenable diseases.

Furthermore, integration with electronic health records and public health systems could enable continuous monitoring of children's health, identifying risks even before birth. The future of neonatal screening is promising, and ML is at the forefront of this revolution.

Conclusion

Neonatal screening with machine learning represents a significant advance in the early detection of rare diseases. With the ability to identify 42 inherited metabolic diseases within 24-48 hours, the system offers a unique opportunity for early intervention, saving lives and preventing severe sequelae. Although ethical and technical challenges remain, the benefits outweigh the risks, and large-scale implementation can transform public health in Brazil and worldwide. The future of preventive medicine begins in the first days of life, and ML is the tool that makes this possible.