Blood that anticipates the future: AI predicts Alzheimer’s four years in advance

While science seeks effective ways to slow the progression of Alzheimer’s, a central problem remains: how to identify early who is at the beginning of the neurodegenerative trajectory? Alzheimer’s disease begins many years before the appearance of clinical symptoms, but traditional diagnostic methods are expensive, invasive, and not very accessible. That is why the search for blood biomarkers — that are sensitive, specific, and easy to measure — has gained strength in recent years.

Many studies have tried to identify these early signs, but few have managed to translate laboratory findings into easy-to-use clinical tools. Brain imaging techniques, such as PET, and cerebrospinal fluid analyses are still the gold standards, but they have obvious limitations: cost, availability, and discomfort. Given this, researchers from the Federal University of Minas Gerais and Kunumi proposed a new predictive model, based on machine learning, that promises to change this scenario. The work, developed by Daniella Castro Araujo and team, revisits an old dream of precision medicine: to predict who will develop Alzheimer’s before the signs appear.

When proteins speak louder than words

The study starts from a powerful and simple idea: blood carries deep information about what is happening in the body, including the brain. The authors' proposal is to build an artificial intelligence model capable of identifying, based on proteins present in plasma, which people with mild cognitive impairment (MCI) will progress to Alzheimer's type dementia within four years.

This question is not trivial. MCI is a borderline condition that can progress to dementia or remain stable for years. Identifying who will progress can allow early interventions, both clinical and behavioral, opening a therapeutic window of opportunity that is not yet fully explored.

By combining data from the ADNI database with machine learning algorithms, the authors went beyond what most previous studies proposed: they built more than one billion different models, exploring multiple combinations among 146 plasma proteins. It is as if they had tested billions of keys to find those that open the lock of Alzheimer’s with greater precision.

An invisible laboratory in every drop of blood

The method proposed by the researchers is, at the same time, radical in scale and pragmatic in result. They used data from 379 individuals with MCI at the ADNI baseline, of whom 176 converted to Alzheimer’s within four years. Plasma protein data were obtained via the Luminex platform, which allows precise measurement of dozens of biomarkers simultaneously.

The chosen algorithm was LightGBM, a powerful machine learning tool based on ensemble decision trees (boosting). To ensure robustness, each model was evaluated with cross-validation in 120 parts, repeated five times — totaling 600 tests per model. A true invisible laboratory running in the background, where each model sought to find hidden risk patterns in the protein data.

The winning model was composed of 12 proteins: ApoB, Calcitonin, C-peptide, CRP, IGFBP-2, IL-3, IL-8, PARC, Serotransferrin, THP, TLSP 1-309 and TN-C. This panel achieved an AUC of 0.91, with an accuracy of 91%, sensitivity of 84% and specificity of 98% to predict the conversion from MCI to Alzheimer’s in up to four years. These numbers are impressive, even more considering that the model is based on a blood sample.

The algorithm that sees the invisible

A thermometer for the future

Among the 12 selected proteins, some stand out for their relation to inflammatory, immunological, and metabolic processes. The most important protein was CRP, known for its association with systemic inflammation. Curiously, lower levels of CRP were associated with a higher risk of conversion, in line with previous studies showing a drop in the levels of this protein in the preclinical phases of Alzheimer’s. The logic may seem counterintuitive: less apparent inflammation, more risk? But perhaps what is at stake here is a failure of the immune system in the initial response.

The puzzle of interactions

Beyond the isolated role of each protein, the model reveals the complexity of the interactions between them. Using the SHAP algorithm, the researchers estimated the individual impact of each protein on risk prediction. Proteins such as IL-8, THP, and PARC also had strong correlations with the prediction, all with lower levels associated with higher risk. By analogy, it is as if the body's alarm system were silently shutting down while the fire begins.

Proteins that scream danger

Serotransferrin and IL-3, on the other hand, showed positive correlations with risk — that is, the higher their levels, the greater the chance of conversion to Alzheimer’s. Serotransferrin, involved in iron transport, has already been associated with oxidative stress processes and beta-amyloid accumulation. IL-3, an immune cytokine, seems to have a neuroprotective role, although some studies also associate it with brain inflammation.

The mathematics of precision

The model's performance is not just rhetorical — it is statistically robust. With an AUC of 0.91 ± 0.01, accuracy of 91%, sensitivity of 84% and specificity of 98%, the model stands out even when compared to cerebrospinal fluid (CSF) biomarkers. In a subgroup of 178 individuals with CSF data, the AI model had a sensitivity of 88% and specificity of 97%, against 73% and 55% of traditional markers. It's like switching from a flashlight to a spotlight.

A visual map of the invisible

The authors also used dimensionality reduction algorithms (t-SNE) to visualize the differences between the groups. When plotted based on protein values, the groups overlapped. But when the data was transformed into SHAP values — which reflect the individual impact of each protein — the separation between converters and non-converters became clear. That is, what matters is not just how much of each protein exists, but how the set behaves in relation to the others.

When doubt is the best trigger for new discoveries

If on one hand the results are promising, on the other, they force us to maintain a critical and questioning stance. Many of the biomarkers identified still do not have clear explanations about their relationship with Alzheimer’s. It is possible that some of them are epiphenomena — side effects of other processes, without a direct causal relationship. This does not diminish their predictive value, but requires caution before translating the model into a clinical tool.

Furthermore, the authors warn of the need for external validation. The model was trained and tested with data from the same database (ADNI), and although cross-validation is rigorous, it does not replace tests in independent populations. It is also not clear how factors such as ethnicity, comorbidities, or medications could influence plasma proteins.

Another important limitation is that some of the proteins used are not yet routinely measured in clinical practice, and their dosage may require sophisticated laboratory infrastructure. This may limit immediate applicability, but also opens opportunities for the development of more accessible and specific tests in the future.

What if we could anticipate Alzheimer’s like we anticipate a storm?

What this study shows us, above all, is that the future of precision medicine lies in the integration between molecular biology and artificial intelligence. By combining a panel of 12 plasma proteins with robust algorithms, the authors were able to predict the progression of Alzheimer’s with very high accuracy and in advance. This not only paves the way for earlier diagnoses but also for preventive and personalized interventions.

And if, in a few years, a simple blood test is enough to know who is about to develop Alzheimer’s? What kind of public policies, clinical strategies, and ethical dilemmas will this generate? Are we prepared to deal with the future when it presents itself so far in advance?

The complete reading of this paper is a provocation for all who wish to rethink how we treat neurodegenerative diseases. More than a technical advance, it invites us to imagine new futures for cognitive health. Worth every line.