Identifying Medicare Beneficiaries With Delirium

Abstract

Background: Each year, thousands of older adults develop delirium, a serious, preventable condition. At present, there is no well-validated method to identify patients with delirium when using Medicare claims data or other large datasets. We developed and assessed the performance of classification algorithms based on longitudinal Medicare administrative data that included International Classification of Diseases, 10th Edition diagnostic codes.

Methods: Using a linked electronic health record (EHR)-Medicare claims dataset, 2 neurologists and 2 psychiatrists performed a standardized review of EHR records between 2016 and 2018 for a stratified random sample of 1002 patients among 40,690 eligible subjects. Reviewers adjudicated delirium status (reference standard) during this 3-year window using a structured protocol. We calculated the probability that each patient had delirium as a function of classification algorithms based on longitudinal Medicare claims data. We compared the performance of various algorithms against the reference standard, computing calibration-in-the-large, calibration slope, and the area-under-receiver-operating-curve using 10-fold cross-validation (CV).

Results: Beneficiaries had a mean age of 75 years, were predominately female (59%), and non-Hispanic Whites (93%); a review of the EHR indicated that 6% of patients had delirium during the 3 years. Although several classification algorithms performed well, a relatively simple model containing counts of delirium-related diagnoses combined with patient age, dementia status, and receipt of antipsychotic medications had the best overall performance [CV- calibration-in-the-large <0.001, CV-slope 0.94, and CV-area under the receiver operating characteristic curve (0.88 95% confidence interval: 0.84-0.91)].

Conclusions: A delirium classification model using Medicare administrative data and International Classification of Diseases, 10th Edition diagnosis codes can identify beneficiaries with delirium in large datasets.

Figure 1.. Sampling Approach

The workflow delineates our sampling procedure to build our final reconstructed sample for analysis, beginning with 40,690 FFS Medicare beneficiaries aged 65-years and older within the ACO. We perform stratified random sampling based on the pretest likelihood based on administrative data. Abbreviations: EHR, electronic health record; ICD-10, International Classification of Diseases, 10th Edition; IPW, inverse proportional weights.

Figure 2.. Calibration Plot (Model 8)

We grouped patients by their deciles of estimated probability and plotted the mean of each group’s expected (predicted) vs. observed (whether they had delirium). If the model were perfectly calibrated, the cross-validated slope would be 1, the calibration-in-the-large would be 0, and all points would lie on the dashed 45-degree line. Abbreviations: CV, cross-validation; CITL, calibration-in-the-large

Figure 3.. Area Under the ROC curve: Tradeoff Between Sensitivity and Specificity

The figure is based on the analytic sample. Area-under-receiver-operating-curve (AUROC) = 0.87. Figure 3 provides the AUROC related to varying sensitivity and 1-specificity and illustrates the point that AUROC is maximized. We display in Figure 3 the performance reference to Model 8 (i.e., best simple model), which used the “Refined” diagnosis codes, plus a count of hospitalizations (inpatient stays) and outpatient visits with delirium diagnoses, count of delirium-associated (antipsychotic) drug fills, and indicators for dementia status and age. Abbreviations: AUROC, area-under-receiver-operating-curve