Spanish AI algorithm detects cancer from blood test, stalled by lack of funding

Isabel Apolonia Yeste Sánchez was studying psychology at a cancer charity in southern Spain when she first encountered the disease up close. She did not imagine then that years later she would teach a machine to recognize it. Pablo José Suárez Pedrajas, a biomedical researcher from Seville who had won awards for his work on multiple myeloma, did not expect to end up calculating market risk at Banco Santander. Yet between them sits a shared project: an artificial intelligence algorithm that can predict the presence of cancer by analyzing just eight proteins in a patient's blood, along with their age. Samsung recognized the work with a national AI award months ago. Today it sits dormant, waiting for data and resources that have not materialized.

Yeste's career pivot came during the pandemic, when the job market froze. She had completed a degree in psychology and worked with cancer patients, then specialized in neuroscience. When lockdowns hit, she made a sharp turn. "I thought, this is the moment to change direction," she recalls. "I could study computer science—everything was going digital." She completed a web development program, discovered an unexpected calling, and won a prize for her first project. That success led her to Samsung Innovation Campus, an artificial intelligence course where she met Suárez virtually. He brought different but complementary training: a degree in basic and experimental biomedicine, recognition from the cancer charity for immunotherapy research, and a master's degree that had introduced him to data analysis, predictive modeling, and clinical data mining. But like Yeste, he had stepped away from academia. "I disconnected from science and the university because I don't identify with the culture there," he says. "The salaries aren't good. People are treated poorly." When Yeste moved temporarily to Seville, they met in person and decided to collaborate on the Samsung course. Suárez proposed the topic: early cancer detection using artificial intelligence.

The numbers driving their work are stark. For lung cancer, five-year survival in advanced stages hovers around five percent. Caught early, it exceeds sixty-five percent. "There is an enormous need to detect cancer in early stages," Suárez explains. Mammography has proven effective for breast cancer, but other tumors—lung cancer chief among them—lack equivalent non-invasive screening methods. Colonoscopies, cervical smears, prostate-specific antigen tests: all exist, but all share a flaw. They are invasive, expensive, or specific to a single cancer type. "Our idea was to create a model that could predict whether a patient might have early-stage cancer, not through invasive tests but through blood," Suárez says.

They began with a dataset published in a high-impact scientific journal. It contained information on patients with and without cancer: concentrations of thirty-nine plasma proteins, an "omega score" estimating the probability of mutations in circulating DNA, plus clinical data like age, sex, and ethnicity. Forty-three variables in total. The challenge was reduction. "Our goal was to make the model as economical as possible by minimizing the number of features needed for prediction," Suárez explains. Using mathematical and statistical methods, they achieved something notable: with only eight plasma proteins and patient age, their model discriminated between cancer presence and absence with very high precision. They surpassed the accuracy of the original study's authors while using fewer variables. "What matters is the pattern that reflects how the disease changes," Suárez clarifies. This is not about measuring a single tumor marker. It is about artificial intelligence identifying patterns in how multiple biomarkers behave together—a logic fundamentally different from traditional tests that analyze parameters in isolation.

Yeste developed the user interface: a simple application where a doctor enters the eight protein concentrations and patient age. The system returns a result: cancer or no cancer. "I didn't want it to stay as code analysis," she explains. "Just a form: you enter the data, press enter, a window pops up: cancer, no cancer." They tested the system by recreating real clinical cases from their dataset. The model predicted correctly. But a model trained on existing data is not yet a clinical tool. "The next step would be clinical validation," Suárez says, "to see how it predicts with data different from what it was trained on." That is where they hit a wall. "The most limiting factor in building a predictive model is data," he says. And data requires money, institutional partnerships, access to hospitals and research centers. They contacted companies and associations. Nothing came through. "We would need collaboration with private institutions, public administrations, or companies," Suárez says. Yeste adds: "We would need data and more support to keep developing this ourselves."

Neither researcher is affiliated with a university. Neither has full-time availability for the project. Yeste currently works for the Andalusian regional government developing a subsidy management platform; Suárez works in market risk at Banco Santander, a field he enjoys for its mathematical dimension. Both acknowledge that developing the project meant "sacrificing sleep." Moving forward would require, they say, "strong backing" and "support." The model currently classifies in binary: cancer or no cancer. A second phase—never developed—would have involved a second algorithm identifying the specific tumor type in positive cases. Why was it not built? "Lack of time and resources," Suárez answers. There is also a technical problem: dividing cases by cancer type drastically reduces the number of instances per category. "You would need unrestricted data access to move forward robustly," he explains. Suárez emphasizes another aspect he considers crucial: understanding why the model fails. "We tried to interpret why the model was wrong, where it was wrong, why it was right, where it was right." Quantifying error is not enough; you must understand it. But that too requires time and data they do not have.

The model is also transferable. Not only to other cancer types but to other diseases: Alzheimer's, ALS, multiple sclerosis. "The model tells you what the defining characteristics of a disease are," Suárez explains. It is, in essence, a mathematical search for biomarkers that should then be validated in the laboratory. "You can adapt it to any disease if you have the data," Yeste summarizes. Their story illustrates a paradox familiar to Spain's research and development system: two trained researchers with validated ideas and institutional recognition—Samsung awarded them in September—but no structure to support the leap from prototype to clinical application. Suárez reflects on university culture: "Universities prioritize publishing; pharmaceutical companies fund more." But pressure to publish in high-impact journals does not always align with implementing practical tools. Yeste holds onto hope: "I would very much like to continue with the project. If there is backing, absolutely." Suárez is more cautious but does not close the door. Both remain open to collaboration. Meanwhile, the algorithm waits on a server, awaiting data to validate what the equations already suggest: a model capable of reading patterns in blood, democratizing early detection, reducing diagnostic costs and time.