ROUTINELY COLLECTED DIGITAL BIOMARKERS FOR PREDICTING PROGRESSION FROM MILD COGNITIVE IMPAIRMENT TO DEMENTIA: A SYSTEMATIC REVIEW
Author(s)
Xiaomo (Shawn) Xiong, MS, PhD1, Eszter Garami, MPharm1, Z. Kevin Lu, PhD2;
1James L Winkle College of Pharmacy, University of Cincinnati, Division of Pharmacy Practice and Administrative Sciences, Cincinnati, OH, USA, 2College of Pharmacy, University of South Carolina, Department of Clinical Pharmacy and Outcomes Sciences, Columbia, SC, USA
1James L Winkle College of Pharmacy, University of Cincinnati, Division of Pharmacy Practice and Administrative Sciences, Cincinnati, OH, USA, 2College of Pharmacy, University of South Carolina, Department of Clinical Pharmacy and Outcomes Sciences, Columbia, SC, USA
OBJECTIVES: Although biomarker- and neuroimaging-based models show promise in predicting dementia risk, digital biomarkers from routinely collected electronic health records (EHRs) or claims offer potential for population-level monitoring. This study evaluated the performance and methodological rigor of models predicting progression from mild cognitive impairment (MCI) to dementia using routinely collected real-world data.
METHODS: We systematically searched PubMed and Embase (from inception to July 1, 2025) for studies developing or validating models predicting progression from MCI to all-cause dementia using routinely collected EHRs or administrative claims. Eligible studies included adults (18 years or older) with MCI at baseline. Two reviewers independently screened studies and extracted data on data source characteristics, predictor domains, modeling approaches, validation methods, and performance metrics.
RESULTS: A total of 2,297 abstracts were screened after removing duplicates, with 72 undergoing full-text review and 7 meeting inclusion criteria. All studies were conducted in the United States, using regional academic or consortium-based EHRs (n = 4), national claims (n = 2), or Veterans Affairs data (n = 1). Sample sizes ranged from 2,525 to 80,138 adults. Most studies (n = 6) focused on progression to Alzheimer’s disease (AD). All studies relied on ICD codes to identify MCI and outcomes. Deep learning models using temporal sequences generally outperformed aggregate-feature models. Specifically, natural language processing (NLP) approaches using unstructured notes (Area Under the Curve [AUC]: 0.88) and structured sequence models (AUC: 0.83-0.87) overperformed traditional and federated models (AUC: 0.67-0.73). Although internal validation was standard, rigorous external validation on independent health systems was limited to two studies.
CONCLUSIONS: AI models using unstructured text and sequences show high potential for predicting MCI-to-AD progression. However, limited external validation, calibration reporting, and reliance on diagnostic codes restrict clinical use. Future research must prioritize multi-site validation and multimodal integration to improve generalizability.
METHODS: We systematically searched PubMed and Embase (from inception to July 1, 2025) for studies developing or validating models predicting progression from MCI to all-cause dementia using routinely collected EHRs or administrative claims. Eligible studies included adults (18 years or older) with MCI at baseline. Two reviewers independently screened studies and extracted data on data source characteristics, predictor domains, modeling approaches, validation methods, and performance metrics.
RESULTS: A total of 2,297 abstracts were screened after removing duplicates, with 72 undergoing full-text review and 7 meeting inclusion criteria. All studies were conducted in the United States, using regional academic or consortium-based EHRs (n = 4), national claims (n = 2), or Veterans Affairs data (n = 1). Sample sizes ranged from 2,525 to 80,138 adults. Most studies (n = 6) focused on progression to Alzheimer’s disease (AD). All studies relied on ICD codes to identify MCI and outcomes. Deep learning models using temporal sequences generally outperformed aggregate-feature models. Specifically, natural language processing (NLP) approaches using unstructured notes (Area Under the Curve [AUC]: 0.88) and structured sequence models (AUC: 0.83-0.87) overperformed traditional and federated models (AUC: 0.67-0.73). Although internal validation was standard, rigorous external validation on independent health systems was limited to two studies.
CONCLUSIONS: AI models using unstructured text and sequences show high potential for predicting MCI-to-AD progression. However, limited external validation, calibration reporting, and reliance on diagnostic codes restrict clinical use. Future research must prioritize multi-site validation and multimodal integration to improve generalizability.
Conference/Value in Health Info
2026-05, ISPOR 2026, Philadelphia, PA, USA
Value in Health, Volume 29, Issue S6
Code
MSR111
Topic
Methodological & Statistical Research
Topic Subcategory
Artificial Intelligence, Machine Learning, Predictive Analytics
Disease
SDC: Neurological Disorders