Leveraging Medicare Claims to Reduce False Positives from 87% to 12% - A Two-Stage Approach to Dementia Screening

Sep 30, 2026

Dementia screening in the United States misses substantial numbers of cases, partly because Medicare’s yearly cognitive screening reaches fewer than a quarter of eligible patients and community tools generate mostly false positives. This poster presents a two-stage strategy that first uses machine learning on Medicare claims to exclude likely healthy individuals, then applies a brief cognitive screener to those flagged as higher risk. Analysis of five years of claims supports reducing the MiniCog false positive rate from 84.3% to 12.7%.

LanguageEnglish

CategoryHealthcare

FormatPDF

Pages2

Reading Time5 min

AccessFree to download

Background

Method

Result

Conclusion

DOI: 10. 1002/alz70856_107487 BIOMARKERS POSTER PRESENTATION BIOMARKERS (NON-NEUROIMAGING) Leveraging medicare claims to reduce false positives from 87% to 12%, a two-stage approach to dementia screening MacKenzie Tweardy Keith J Yoder Spencer Gerrol Ché Lucero SPARK Neuro, Inc., New York, NY, USA Correspondence MacKenzie Tweardy, SPARK Neuro, Inc., New York, NY, USA. [Email:](Email: mackenzie.tweardy@sparkneuro.com)[ mackenzie.tweardy@sparkneuro.com](Email: mackenzie.tweardy@sparkneuro.com) Abstract Background: Some 40% - 60% of dementia goes undiagnosed in the United States, creating missed opportunities and worsening outcomes for patients. Medicare recommends yearly cognitive screening for the elderly, but fewer than a quarter of those patients get cognitive assessments. Why might this be? Community screening for dementia with common pen-and-paper instruments yields mostly false positives (around 85%). Here, we demonstrate that these false positives can be dramatically reduced by first using a machine-learning model trained on Medicare claims data to exclude individuals most likely to be healthy to focus assessments on those most likely to have undiagnosed dementia. Method: We analyzed a 5-year span of Medicare claims data, reflecting 40 million claims from1.9million individuals. We extracted features by calculating the proportion of claims within a given time period that were related to a set of literature-based diagnoses and risk factors, such as diabetes and hypertension. We then trained a gradient-boosted decision tree (XGBoost) to detect undiagnosed dementia. To understand whether this model could be combined with a cognitive screener to reduce false positives, we evaluated its expected performance at detecting undiagnosed dementia in a hypothetical population of 100,000 older adults. We compared the false positive rate for detecting new dementia cases when a brief cognitive screener was used alone or applied only to those first flagged by the claims-based model. Result: Using performance from a recent meta-analysis of sensitivity (76%) and specificity (83%), and an assumed undiagnosed dementia prevalence of 4%, the MiniCog resulted in a false positive rate of 84.3%. In contrast, administering the Mini-Cog only to those individuals flagged by the claims-based model, reduced the false positive rate to 12.7%. This approach decreased the total number of Mini-Cog assessments from 100,000 to 2,924 while identifying 1,347 new dementia cases. Conclusion: Reducing false positives is key to making screening practical. We find that using a machine learning approach to claim-analysis can identify patients who should This is an open access article under the terms of the Creative Commons Attribution License, which permits use, distribution and reproduction in any medium, provided the original work is properly cited. © 2025 The Alzheimer’s Association. Alzheimer’s & Dementia published by Wiley Periodicals LLC on behalf of Alzheimer’s Association. Alzheimer’s Dement. 2025;21(Suppl. 2):e107487. [https://doi.org/10.1002/alz70856_107487](https://doi.org/10.1002/alz70856_107487) wi[leyonlinelibrary.com/journal/alz](leyonlinelibrary.com/journal/alz) 1of2 be screened with a test like the Mini-Cog. This creates a practical path to discovering and diagnosing dementia and connecting patients with appropriate care.

FAQs

Why do standard dementia community screenings produce many false positives?

Community screening using common pen-and-paper instruments yields mostly false positives, around 85%, which reduces practical usefulness and can divert resources away from those with undiagnosed dementia.

What is the two-stage approach proposed in the study?

First, a gradient-boosted decision tree is trained on Medicare claims to identify individuals most likely to have undiagnosed dementia. Second, a brief cognitive screener (MiniCog) is administered only to people flagged by the claims-based model.

How much can false positives be reduced using the two-stage strategy?

With MiniCog used alone, the false positive rate is estimated at 84.3%. Restricting MiniCog to individuals flagged by the claims-based model reduces the false positive rate to 12.7%.

Sep 30, 2026

Uploaded by CatatanPagi