Brazil is living a moment of enthusiasm with artificial intelligence in healthcare. Startups launch risk prediction models, hospitals announce clinical decision-support algorithms, insurers test population stratification tools, and the regulatory discussion about AI use in healthcare gains ground at ANS, the Federal Council of Medicine, and the National Data Protection Authority. The sophistication of the models grows every year. The public debate concentrates on them: their accuracy, their biases, their regulation, their safety.
There is a paradox at the center of this enthusiasm that rarely enters the conversation. The overwhelming majority of the clinical data these sophisticated models consume still arrives in formats that weren't made to be consumed by machine: PDF, free text, report image, non-standardized spreadsheet. The bottleneck of AI advancement in Brazilian healthcare isn't in the models. It is in the data layer that feeds them.
The model is only as good as the data it receives
The quality of any artificial intelligence model is limited by the quality of the data it operates on. This is one of the most established propositions in the field, and it applies with particular force to healthcare. Shortliffe and Cimino, in a central reference of biomedical informatics, observe that data quality is the foundation on which all health automation is built, and that unstructured data compromises that foundation at the source.¹
In Brazilian practice, that foundation is fragile. A classic study estimated that around 80% of health data is in unstructured formats, hindering the systematic extraction of clinical insights.² When a risk prediction model needs to consume a patient's laboratory history, and that history exists as a collection of PDFs from different laboratories, with distinct nomenclatures, varied units, and proprietary reference ranges, the model doesn't receive data. It receives documents that need to be converted into data before any inference.
That conversion, known in technical jargon as the preprocessing pipeline, is where the paradox materializes. Each time the model needs to process a patient's history, the system runs OCR to extract text from the PDFs, parsing to identify values, mapping to recognize biomarkers, unit normalization, and an attempt at reference range harmonization. Each of these steps introduces opportunities for error and fidelity loss. The sophisticated model runs, but it runs over a layer reconstructed at each inference, with computational cost and quality degradation that the discourse about the model's sophistication ignores.
The invisible cost of permanent reconstruction
When the data layer isn't structured at the source, the structuring work doesn't disappear. It shifts into the model's pipeline, where it has to be redone repeatedly.
This has three consequences that compromise the real value of AI in healthcare. The first is computational cost. Reconstructing the data structure at each inference, instead of consuming already-structured data, multiplies the necessary processing and the associated cost. At scale, this means a substantial part of the computational resources of a healthcare AI operation is being spent not on inference, but on converting documents that could have been structured a single time.
The second consequence is fidelity loss. Each step of converting a PDF into structured data is an opportunity for error: a value poorly extracted by OCR, a biomarker mapped incorrectly, a unit not converted. These errors propagate into the model's inference, and the model has no way to distinguish a conversion error from real clinical data. The model's announced accuracy, measured in controlled conditions with clean data, doesn't hold when the model operates over data reconstructed from PDFs of variable quality.
The third consequence is the impossibility of auditing. Healthcare AI models are increasingly subject to explainability and traceability requirements, both for clinical and regulatory reasons. When the data layer feeding the model is reconstructed at each inference from unstructured documents, tracing why the model produced a given output becomes extraordinarily difficult, because the input data itself doesn't have stable, auditable provenance.
Why the debate is in the wrong place
The public debate about AI in healthcare in Brazil concentrates on the models because the models are visible, impressive, and easy to discuss. The data layer is invisible, tedious, and hard to turn into a headline. But that asymmetry of attention produces a misallocation of effort and investment.
Investing in increasingly sophisticated models while keeping the data layer in PDF and free text is building additional floors over a foundation that hasn't been reinforced. The marginal gain of a more sophisticated model is limited by the quality of the data it receives, and when that quality is low, the more sophisticated model doesn't deliver proportionally more value than the previous one. The return on investment in model sophistication decreases rapidly when the data layer remains the bottleneck.
The literature on health interoperability reinforces that the semantic structuring of data, and not just its availability, is what enables its advanced analytical use.³ Standards like LOINC for laboratory test identification were developed over decades precisely to solve the problem of making clinical data comparable and machine-processable.⁴ The healthcare AI that ignores this infrastructure and tries to extract meaning directly from PDFs is wasting the standardization work the global ecosystem has already done.
Structuring at the source as the condition for AI to work
Correcting the paradox doesn't go through better models. It goes through structuring data at the source, so that the model consumes genuinely structured data instead of reconstructing it at each inference. When a patient's laboratory history exists as a harmonized time series, with biomarkers mapped to recognized standards, normalized units, and documented reference ranges, the AI model receives exactly what it needs to operate with the accuracy its sophistication promises.
This structuring has a multiplier effect. The same harmonized data layer that feeds a risk prediction model also feeds the other models, the management dashboards, the clinical alerts, and the population analyses. Structuring once, at the source, replaces the repeated reconstruction inside each pipeline. Computational cost falls, fidelity rises, auditing becomes possible, and the investment in model sophistication starts generating the return that data quality previously limited.
The regulatory context reinforces the urgency. Ordinance GM No. 8,276, published by the Ministry of Health in October 2025, established the Laboratory Test Result Information Model within the National Health Data Network, requiring laboratory results to be structured with recognized terminologies and sent to RNDS.⁵ The direction is clear: structuring laboratory data is ceasing to be optional. Institutions that structure their data at the source will simultaneously be complying with regulation and building the foundation on which healthcare AI actually works.
This is where OpenHealth Technologies operates. The platform automatically correlates multiple data streams with rigorously validated logical layers of laboratory tests, transforming laboratory data in any format, including PDF and free text, into structured data, mapped to LOINC, comparable and machine-processable, with over 3,500 biomarkers covered. For healthcare AI initiatives, this means the model stops reconstructing the data layer at each inference and starts consuming data structured at the source, with the computational cost, fidelity, and auditability that the model's sophistication requires to deliver the value it promises.
Learn how your institution can build the structured data layer that artificial intelligence in healthcare needs to work as it should, instead of reconstructing it at each inference.