AI model predicts which patients with memory problems will develop dementia within two years

A team of researchers in Florida has built an artificial intelligence system that can predict, with remarkable accuracy, whether a person diagnosed with mild cognitive impairment will slide into dementia within the next two years. The study, published in the journal GeroScience, analyzed data from hundreds of older adults enrolled in the Alzheimer’s Disease Neuroimaging Initiative and demonstrated that a single visit to the clinic could provide enough information to forecast who will remain stable and who will deteriorate. The work arrives at a pivotal moment, as new Alzheimer’s therapies increasingly depend on identifying patients early, while the disease process is still unfolding rather than after irreversible damage has accumulated.

Mild cognitive impairment, often abbreviated as MCI, is one of the most diagnostically frustrating states in medicine. It describes a genuine decline in memory or thinking that goes beyond normal aging, yet it sits on a fault line: some people with MCI stabilize or even revert to normal cognition, while others progress to full dementia, most commonly Alzheimer’s disease. Previous meta-analyses of dozens of cohort studies have shown that annual conversion rates vary enormously between clinical settings, and a large fraction of people diagnosed with MCI never develop dementia at all. This heterogeneity is precisely why clinicians have long sought reliable tools to separate the two trajectories, and why misclassification carries such high stakes, either exposing stable patients to unnecessary treatment or delaying care for those truly at risk.

The research team, led by Bipul Simkhada of Florida International University together with colleagues at the University of Miami, the University of Florida, and Clemson University, tackled the problem with a multimodal machine-learning framework. They assembled 2008 samples drawn from 828 unique individuals with MCI in the Alzheimer’s Disease Neuroimaging Initiative, one of the largest and most thoroughly characterized longitudinal dementia datasets in existence. Each sample combined three distinct families of information: cognitive and functional assessments gathered during clinical visits, demographic and genetic variables including APOE genotype, and structural features extracted from magnetic resonance imaging of the brain. Based on what actually happened over a two-year follow-up window, subjects were labeled as either stable MCI, comprising 1306 samples, or progressive MCI, comprising 702 samples that converted to dementia.

Methodological rigor was central to the design. Rather than relying on a single train-test split, which can inflate performance estimates, the researchers used nested cross-validation, a scheme in which model selection and performance evaluation are separated at every step. They also addressed the inherent class imbalance between stable and progressive samples using the synthetic minority over-sampling technique, known as SMOTE, and paid careful attention to feature scaling, calibration of predicted probabilities, and the selection of clinically meaningful decision thresholds. These details matter because a prediction tool that reports probabilities poorly calibrated against real-world outcomes can mislead physicians even when its raw accuracy appears impressive. The team’s final model was a calibrated version of XGBoost, a gradient-boosted decision tree algorithm widely used in biomedical prediction for its ability to capture nonlinear interactions between variables without requiring enormous computing resources.

The results were striking. During nested cross-validation, the multimodal XGBoost model achieved a balanced accuracy of 81.28 percent, with a standard deviation of 2.56 percent, meaning that it correctly identified both converters and non-converters at nearly equal rates despite the unequal group sizes. On an independent held-out test set that the model had never seen during training, balanced accuracy reached 81.99 percent, with a bootstrap mean of 81.97 percent and a 95 percent confidence interval spanning 78.22 to 85.51 percent. Crucially, the multimodal combination outperformed any single data type used alone. Cognitive tests, demographic risk factors, and brain imaging each carried predictive signal, but their fusion proved strictly superior, underscoring a complementary relationship between the observable decline in day-to-day function and the silent anatomical erosion happening inside the brain.

What elevates the study beyond a simple accuracy benchmark is its insistence on explainability. Black-box predictions have historically been a barrier to clinical adoption, so the researchers interrogated their model using SHAP, or SHapley Additive exPlanations, a technique derived from cooperative game theory that assigns each input feature a precise contribution to every individual prediction, alongside permutation feature importance, which measures how much performance degrades when a given variable is shuffled. The analyses converged on a coherent and clinically intuitive picture. The strongest predictors included functional impairment measures, specifically the Functional Activities Questionnaire and the Clinical Dementia Rating Sum of Boxes, which capture whether a person is struggling with finances, medications, shopping, and other instrumental tasks of daily living.

Cognitive performance scores followed closely, particularly the 11-item and 13-item versions of the Alzheimer’s Disease Assessment Scale cognitive subscale and the Mental State Examination, familiar bedside instruments that quantify memory, orientation, language, and attention. Alongside these behavioral markers, the model leaned heavily on biological signals: the APOE ε4 allele, the best-established genetic risk factor for sporadic Alzheimer’s disease, and a constellation of structural brain abnormalities. The imaging features that mattered most mapped onto regions long implicated in the disease, including the hippocampus, the memory-forming structure that atrophies earliest, the lateral ventricles that expand as surrounding tissue shrinks, the parietal and temporal lobes where cortical thinning tracks disease spread, and the amygdala, a deep structure involved in emotion and memory circuits.

The convergence of these strands is perhaps the study’s most important conceptual message. Neither cognitive test scores nor genetic risk nor imaging atrophy alone tells the whole story, but when a patient shows functional decline on questionnaires, poor performance on structured cognitive batteries, carries an ε4 allele, and exhibits measurable shrinkage in exactly the brain regions vulnerable to Alzheimer’s pathology, the probability of near-term conversion rises sharply. The researchers describe this alignment of cognitive-functional impairment and region-specific neurodegeneration as the strongest indicator of impending dementia, a finding that reinforces the biological intuition that clinical symptoms and structural damage advance together during the transition period.

The practical implications are considerable. Because all of the input data can be collected during a single clinical visit, without longitudinal repeated scans or experimental fluid biomarkers, the framework could in principle be deployed in memory clinics using instruments that clinicians already administer. An accurate two-year risk estimate would allow physicians to triage patients, directing high-risk individuals toward early anti-amyloid therapy, closer monitoring, and participation in clinical trials, while sparing lower-risk patients unnecessary interventions and anxiety. It would also sharpen the design of prevention studies, which currently spend enormous resources enrolling participants who never progress. The team has made its analysis code publicly available on GitHub, and the underlying data come from the openly accessible ADNI database, allowing other groups to validate and extend the approach.

Challenges remain before such models enter routine practice. ADNI participants are volunteers who tend to be well educated and predominantly of European ancestry, and recent systematic reviews have flagged limited generalizability and high risk of bias across many published MCI prediction models, so independent validation in diverse, real-world populations will be essential. Still, the combination of robust validation methodology, calibrated probabilities, explainable outputs, and accuracy above 80 percent using only routine clinical data marks a meaningful advance. As the population ages and the number of people living with mild cognitive impairment climbs into the tens of millions worldwide, tools that can distinguish, at the first appointment, who will decline and who will not, may fundamentally reshape how medicine confronts the earliest and most treatable stage of dementia.

Subject of Research: Multimodal machine-learning prediction of two-year conversion from mild cognitive impairment to dementia

Article Title: Multimodal prediction of MCI-to-dementia conversion over a two-year window

Article References: Simkhada, B., Liang, T. Y., Cui, X., Adeyosoye, M., Simkhada, B., Cabrerizo, M., Cid, R. C., Burke, S. L., Barreto, A., Rishe, N., Loewenstein, D. A., & Adjouadi, M. (2026). Multimodal prediction of MCI-to-dementia conversion over a two-year window. GeroScience. https://doi.org/10.1007/s11357-026-02551-x

Image Credits: AI Generated

DOI: 10.1007/s11357-026-02551-x

Keywords: mild cognitive impairment, Alzheimer's disease, dementia, machine learning, XGBoost, explainable AI, SHAP, structural MRI, APOE ε4, hippocampal atrophy, ADNI, GeroScience