AI against breast cancer: the challenge is no longer the algorithm, but diverse data, validation and privacy

By IFM Noticias · October 20, 2026.

Fran Ernesto Romero Álvarez, director of the Data Engineering and Artificial Intelligence Program at the Universidad Católica de Colombia, explained during Breast Cancer Prevention Month how these technologies are applied in clinical settings. According to the expert, the models cross-reference clinical, genetic, lifestyle and imaging data to identify high-risk profiles and prioritize the most urgent cases.

The most significant nuance is the shift from diagnosis to prediction. Romero Álvarez notes that applications have emerged that estimate the risk of developing cancer from mammograms. In other words, they don't just detect a tumor that is already present; they assess the likelihood that one will appear.

According to the article, four types of data feed these systems: clinical and family history, habits (alcohol, tobacco, physical activity), genetic tests and imaging (mammograms, ultrasounds and MRIs). The quality of the prediction depends on the quality and diversity of that data.

The expert warns of specific risks: diagnostic errors, bias and privacy breaches. The article explains that a model without rigorous validation can produce false positives (flagging a disease that doesn't exist) or false negatives (missing one that does). If a population is underrepresented in the training data, the system may be less accurate for it. It also mentions re-identification: techniques that can reverse patient anonymization.

As a next step, he points to large multimodal models (LMMs). Unlike language models, which process text, they integrate text, images and video. In breast cancer, a tool could correlate the clinical note, the mammogram and genetic markers at the same time.

Our take: the article is aimed at a general audience and provides no accuracy figures, clinical trials or names of specific tools. It should therefore be read as a map of opportunities and risks, not as evidence of results. What is solid is the underlying diagnosis: in medicine, AI's bottleneck is rarely the model, but the data. Clean, integrated data that is representative of diverse populations is the true critical infrastructure.

That point has a direct consequence for Latin America. In general, large medical models are trained on data from wealthy countries, so their performance in other populations is an unknown until validated locally. Whoever invests in their own well-governed health data will have more reliable tools and greater sovereignty. That is also where the professional opportunity the expert highlights lies for data engineers and AI specialists.

In the short term, the transition will be uneven: poorly validated systems can cause anxiety, overdiagnosis or overconfidence, and hospitals with fewer resources will take longer to benefit. In the long term, the outlook is hopeful. Earlier detection and personalized risk assessment are concrete steps toward a medicine that prevents rather than reacts, and toward the broader goal of eradicating diseases like this one. But that future is built through constant validation, evidence-based regulation and respect for privacy, not through enthusiasm.