Objective:

To develop automated screening tools that identify patients with treatable causes of cognitive impairment from routine clinical documentation. 

Background:

Specialist evaluation is often called on to distinguish reversible from neurodegenerative cognitive impairment, with frequent delays before identifying treatable causes. Patient-reported outcome measures and other data documented during routine care may facilitate automated screening. 

Design/Methods:

Data from 863 patients at an academic memory center were analyzed. Validated instruments (Patient Health Questionnaire-9 (PHQ-9), Generalized Anxiety Disorder-7 (GAD-7), Epworth Sleepiness Scale, Patient-Reported Outcomes Measurement Information System (PROMIS) pain, medication inventory, and other indicators) were aggregated into composite risk scores quantifying psychiatric burden, chronic pain, sleep disorders, and polypharmacy. Three machine learning algorithms (logistic regression, random forest, gradient boosting) were trained to replicate validated specialist judgments. Specialist diagnostic formulations predicting progression likelihood were validated against longitudinal cognitive trajectories. 

Results:

32.8% of males and 28.7% of females demonstrated treatable rather than progressive impairment.  Non-progressive cases associated with worse mood (z=0.96), pain (z=0.88), and medication risks (z=1.68) compared to progressive diagnoses. Post-concussive symptoms were more common in men (12.7% versus 1.3%, p<0.0001) while functional/psychiatric symptoms were more common in women (3.2% versus 0.9%, p<0.01) with cognitive concerns. Gradient boosting models based on composite risk scores correlated with specialist predictions (AUROC 0.885 for mild impairment, 0.749 for severe impairment). Longitudinal follow-up confirmed prediction accuracy: patients classified as high risk declined 0.71 MoCA points annually (p=0.020) versus those classified as low risk. 

Conclusions:

Automated classification of routine clinical measures correlates with prospectively validated specialist prognostication, potentially supporting future point of care identification of reversible cognitive impairment using clinical documentation.