Leveraging medicare claims to reduce false positives from 87% to 12%, a two‐stage approach to dementia screening

TL;DR

A machine learning model trained on Medicare claims data can reduce false positives in dementia screening from 87% to 12% by focusing assessments on those most likely to have undiagnosed dementia.

Contribution

A novel two-stage dementia screening approach using Medicare claims data and machine learning to drastically reduce false positives.

Findings

Using a machine learning model on Medicare claims reduced false positives from 84.3% to 12.7%.

The approach reduced the number of cognitive assessments needed from 100,000 to 2,924 while identifying 1,347 new dementia cases.

Abstract

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.

We analyzed a 5‐year span of Medicare claims data, reflecting 40 million claims from 1.9 million 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.

Using performance from a recent meta‐analysis of sensitivity (76%) and specificity (83%), and an assumed undiagnosed dementia prevalence of 4%, the Mini‐Cog 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.

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 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.