Nigerian researcher proposes safer path for AI-driven disease prediction

A Nigerian biostatistician, Abdulazeez Alabi, has led a major international study offering health systems a practical roadmap for deploying more reliable and trustworthy risk prediction models for chronic diseases.

Chronic illnesses such as diabetes, heart disease, kidney disease, and chronic obstructive pulmonary disease continue to strain health systems globally, particularly in Africa, where access to advanced care remains uneven, and resources are limited.

In a paper titled “From Logistic Regression to Foundation Models: Factors Associated with Improved Forecasts,” published in the Cureus Journal of Medical Science, a part of Springer Nature, Alabi and his colleagues examined how modern artificial intelligence tools compare with traditional statistical methods in predicting patient outcomes.

Alabi, who is affiliated with the Department of Mathematics and Statistics at Georgia State University, worked alongside Olajide Akinpeloye, Osayimwense Izinyon, Tope Amusa, and Akinwale Famotire.

The research sought to answer a pressing question in modern medicine: when do complex machine learning models outperform traditional approaches, and when does their added complexity introduce new risks in real-world clinical settings?

According to the study, advanced models such as gradient-boosted trees, including XGBoost and LightGBM, “consistently outperformed traditional logistic regression in prediction accuracy,” recording noticeable gains across multiple studies.

However, Alabi warned that these gains come with risks if not properly managed. “Many of these models require post-hoc recalibration before they can be safely used clinically,” he said, noting that without such adjustments, some systems significantly overestimate patient risk.

The study also examined emerging foundation models trained on vast health datasets and found that they hold promise, particularly in low-resource settings. One finding showed that such models could match or exceed traditional systems “using fewer than one percent of the target labels,” a development seen as crucial for hospitals with limited data.

Despite the growing enthusiasm for artificial intelligence, the researchers found that logistic regression remains highly relevant. Alabi noted that “when patient data relationships are stable and linear, logistic regression can match or even exceed more complex models without the added cost and expertise.”

A key contribution of the research is its emphasis on what the team described as a “calibration-first approach.” Alabi explained that health systems have focused too heavily on accuracy while neglecting whether predictions reflect real-world outcomes, warning that “a model that predicts a 70 per cent risk should correspond to actual outcomes, otherwise it leads to poor clinical decisions.”

The researchers proposed clear benchmarks, including a calibration slope between 0.90 and 1.10 and strict performance thresholds, insisting that any model failing to meet these standards “should not be deployed regardless of how impressive its accuracy appears.”

Alabi added that the findings are particularly relevant for Nigeria and other African countries, where healthcare systems are increasingly adopting AI tools. He concluded that while simpler models remain effective in many settings, “foundation models offer a future where even under-resourced hospitals can benefit, provided recalibration and transparency become standard practice.”