Automated identification of older adults at risk for cognitive decline

Abstract

Introduction: Automated models that predict cognitive risk in older adults can aid decisions about which patients to screen in busy primary care settings.

Methods: In this retrospective prediction model development study, we conducted formal cognitive testing on 337 older primary care patients to establish cognitive status. We used up to 5 years of prior discrete-field electronic health record (EHR) data to develop a multivariable prediction model that differentiates patients with impaired versus intact cognition.

Results: The final model included seven easily extractable variables with known associations to cognitive decline: age, race, pulse, systolic blood pressure, non-steroidal anti-inflammatory use, history of mood disorder, and family history of neurological disease. The model demonstrated good discrimination of cognitive status (concordance statistic = 0.72).

Discussion: The cognitive risk model may be useful clinically to prompt for objective cognitive screening in high-risk patients. The use of common, discrete variables ensures relative ease of implementation in EHRs.

Highlights: 337 older primary care patients completed full neuropsychological assessment.Risk modeling used data available in a typical primary care record.The model successfully differentiated patients with/without cognitive impairment.This EHR model offers a passive workflow to identify patients at cognitive risk.

Keywords: decision support; dementia; early diagnosis; electronic health record; machine learning; mild cognitive impairment; prediction model.

FIGURE 1

Model calibration curve. A perfectly calibrated model will produce the ideal 45° line, indicating 100% concordance between predicted (i.e., model‐assigned) probabilities and observed probabilities. The dark line represents the curve from our model, and the shaded regions are 95% confidence bands. The curve follows the ideal line closely, with minor deviations at the extreme ends; this phenomenon is normal in samples where most assigned probabilities fall in the middle of the range. Deviations indicate a slight (2% to 4% points) underestimation of risk for those assigned probabilities of 0.18 or lower and those assigned probabilities of 0.6 or higher. Confidence bands fully encompass the ideal line.