Alzheimer's: The Algorithm That Distinguishes Risk Between Men and Women

When it strikes—approximately 800,000 people in Italy—it doesn't affect everyone equally. Indeed, the numbers show that women suffer from Alzheimer's disease at a higher rate: from 0.7% among those aged 65-69 to 23.6% for women over 90, compared to men whose rates range from 0.6% to 17.6%, respectively. But the way the disease manifests itself also makes a difference: women often experience more rapid cognitive decline and greater impairment of episodic memory. This appears to be due—though research is still investigating—to the role of sex hormones and differences in brain structure. Yet the tools used to diagnose the disease, from tests to more invasive and expensive methods, such as MRI or cerebrospinal fluid analysis, don't sufficiently account for these differences. Now an algorithm could make a difference.

A bias to overcome

"Many neuropsychological tests have historically been developed on predominantly male samples," explains Daniele Caligiore, research director at the Institute of Cognitive Sciences and Technologies (ISTC) of the National Research Council (CNR). "The problem is that these tests are then applied to people of both sexes, but men and women may respond differently. This can lead to diagnostic errors: for example, a woman may score lower on a test, but that score could still indicate good performance if assessed with a scale calibrated for the female sex."

Machine learning for early diagnosis

In this context, machine learning can represent an opportunity, as demonstrated by the project coordinated by Caligiore's research center, the results of which were published in the Journal of Neurological Sciences. "Our goal," the scholar continues, "is to understand which factors should be considered from a gender perspective to aid physicians in early diagnosis, especially for complex diseases like Alzheimer's and Parkinson's, which are closely related."

An Italian collaboration

The work is the result of a collaboration between the CNR (National Research Council), the Milan 4 Research Area, the Mondino Foundation, the University of Pavia, the Santa Lucia Foundation, Sapienza University of Rome, and the start-up AI2Life. The core of the project is a machine learning algorithm capable of predicting and differentiating the onset of Alzheimer's disease based on a patient's sex, using non-invasively collected data such as neuropsychological test scores and sociodemographic information. The algorithm was trained with data from two large international databases, including that of the Michael J. Fox Association. Unlike traditional approaches, which treat data homogeneously, the team trained the system separately on data from men and women. In this sense, the machine learning model not only allows us to predict the probability of developing the disease within a specific time frame (one to five years), but also—thanks to the use of explainable AI that makes the algorithm's decision-making process transparent—to identify which tests are most predictive for each gender.

Test for women and men

Analysis of the results shows that some neuropsychological tests have different predictive value depending on gender. "Machine learning allows us to analyze relative differences in tests and combine multiple parameters," Caligiore continues. "For example, we saw that the Mini-Mental State Examination test is more effective in predicting Alzheimer's disease in women, as is the test that assesses long-term episodic memory (LDELTOTAL), while the test of verbal short-term memory (AVTOT) is more relevant for men. In other words, the system is able to say: 'To better diagnose, test X is important in men, test Y in women.' This is a breakthrough in personalized diagnosis." Education level, and therefore cognitive reserve, was also found to be a determining factor, particularly for women.

A personalized and accessible diagnosis

A distinctive element of the project is the creation of a graphical interface called EMA (ExplAIn Medical Analysis), which allows doctors to directly use the system. Simply enter neuropsychological test scores, and the algorithm provides a risk assessment with a numerical probability. "In the future, we imagine a simple system where patients are administered questionnaires, scores are collected, and the clinician enters them into the interface," Caligiore explains. "The system returns a number: for example, '75% probability of developing Alzheimer's within three years.' It's a predictive tool, useful even for subjects who don't yet show obvious symptoms." Often, he emphasizes, AI research stops at the laboratory stage. Instead, they want to make this tool usable in everyday clinical practice, so that Alzheimer's diagnosis can be more timely, more equitable, and less invasive.

Training with data from Italian patients

The team is now working on a new development phase, based on Italian clinical data. This step is crucial to reduce cultural bias and adapt the system to the characteristics of the European population. "North American data is very useful, but it reflects a lifestyle different from ours," Caligiore concludes. "That's why we're now validating the algorithm with Italian data, to increase accuracy and build a truly effective system in our healthcare context as well."