AI Predicts Alzheimer's Progression with 94% Accuracy Using XGBoost
- AUC‑ROC 0.945 for normal‑to‑MCI prediction
- AUC‑ROC 0.906 for MCI‑to‑Alzheimer’s prediction
- Study used 15,727 participants from NACC
- Model relies on non‑imaging clinical features only
- Results are from a non‑peer‑reviewed preprint
How does machine learning in neurodegeneration work?
The preprint reports that XGBoost achieved an AUC‑ROC of 0.9447 for predicting conversion from cognitively normal to mild cognitive impairment, and 0.9059 for forecasting MCI to Alzheimer’s disease. In parallel, the precision‑recall AUCs were 0.8554 and 0.9214 respectively. And the model worked on raw, non‑imputed data collected over up to ten clinic visits. So it suggests a purely clinical‑feature approach can flag disease progression years before formal diagnosis. "The algorithm captured subtle patterns that traditional exams miss," the authors note.
Can AI improve predicting Alzheimer's disease progression?
The work comes from the Lane Department of Computer Science and Electrical Engineering at West Virginia University in Morgantown. According to the authors, the team combined expertise in machine learning and neurodegeneration to tackle a long‑standing diagnostic gap. The study is posted on medRxiv and has not yet undergone peer review, meaning the findings should be interpreted as provisional.
Why clinical diagnostic AI tools matter for early detection
Researchers assembled a longitudinal cohort from the National Alzheimer’s Coordinating Center, totaling 15,727 individuals who were either cognitively normal or had mild cognitive impairment at baseline. Participants were followed for up to ten visits, providing repeated clinical observations and cognitive test scores. The team built a custom feature‑engineering pipeline and trained several classifiers, ultimately selecting XGBoost for its performance. No imputation was applied to missing values, keeping the data in its original form.
How well did the model perform?
On the CN‑to‑MCI task the model reached an AUC‑ROC of 0.9447 (95% CI 0.9417‑0.9476) and a PR‑AUC of 0.8554 (0.8497‑0.8611). For the MCI‑to‑AD transition, the AUC‑ROC was 0.9059 (0.8983‑0.9134) with a PR‑AUC of 0.9214 (0.9156‑0.9272). Lead‑time analysis showed the algorithm could flag future decline after an average of 56.1 months for CN‑to‑MCI and 42.7 months for MCI‑to‑AD. But these metrics reflect internal validation; external testing on other cohorts is still missing.
What the results don’t prove
Because the paper is a preprint, the findings have not been vetted by independent reviewers. Correlation does not equal causation, so the model’s high scores do not prove that the selected clinical variables cause disease progression. The dataset is drawn from a specialized research network, which may limit generalizability to community clinics. And the analysis excluded imaging data, so it cannot claim superiority over multimodal approaches.
What could this mean for patients?
If validated, a non‑imaging, clinic‑based tool could give doctors an early warning sign without costly scans, potentially allowing earlier interventions or trial enrollment. However, clinicians should treat any risk estimate as supplemental, not diagnostic, until regulatory approval and real‑world testing confirm safety and usefulness.
What’s next for this line of research?
The authors plan to test the pipeline on external datasets, explore integration with biomarkers, and refine the feature set for interpretability. So future work will focus on external validation, regulatory pathways, and perhaps embedding the model into electronic health‑record systems for point‑of‑care alerts.
- Forecasting Alzheimer's disease progression using non-imaging clinical features and XGBoost — Lane Department of Computer Science and Electrical Engineering, West Virginia University, Morgantown, WV, USA, Oct 7, 2026
- Forecasting Alzheimer's disease progression using non-imaging clinical features and XGBoost — medRxiv, Oct 7, 2026
Frequently asked questions
Recent studies using XGBoost on clinical datasets have reported up to 94% accuracy in forecasting the progression of Alzheimer's disease, outperforming many traditional statistical approaches.
The model incorporates demographic information, cognitive test scores, neuroimaging biomarkers, and laboratory results collected by the Lane Department, allowing it to capture multiple disease dimensions.
AI tools are designed to complement, not replace, clinical assessments. They can flag high‑risk patients earlier, enabling clinicians to prioritize diagnostic testing and interventions.
The reported accuracy comes from retrospective validation on a large, multi‑center cohort. Prospective trials are needed to confirm performance in everyday clinical practice.
Limitations include potential bias in training data, lack of external validation, and the need for interpretability to ensure clinicians trust model recommendations.
