Prognosis of Individual-Level Mobility and Self-Care Stroke Recovery During Inpatient Rehabilitation, Part 1: A...

To demonstrate feasibility of generating predictive short-term individual trajectory recovery models after acute stroke by extracting clinical data from an electronic medical record (EMR) system.

Single-group retrospective patient cohort design.

Stroke rehabilitation unit at an independent inpatient rehabilitation facility (IRF).

Cohort of 1408 inpatients with acute ischemic or hemorrhagic stroke with a mean ± SD age of 66 (14.5) years admitted between April 2014 and October 2019 (N=1408).

Not applicable.

0-100 Rasch-scaled Functional Independence Measure (FIM) Mobility and Self-Care subscales.

Unconditional models were best-fit on FIM Mobility and Self-Care subscales by spline fixed-effect functions with knots at weeks 1 and 2, and random effects on the baseline (FIM 0-100 Rasch score at IRF admission), initial rate (slope at time zero), and second knot (change in slope pre-to-post week 2) parameters. The final Mobility multivariable model had intercept associations with Private/Other Insurance, Ischemic Stroke, Serum Albumin, Motricity Index Lower Extremity, and FIM Cognition; and initial slope associations with Ischemic Stroke, Private/Other and Medicaid Insurance, and FIM Cognition. The final Self-Care multivariable model had intercept associations with Private/Other Insurance, Ischemic Stroke, Living with One or More persons, Serum Albumin, and FIM Cognition; and initial slope associations with Ischemic Stroke, Private/Other and Medicaid Insurance, and FIM Cognition. Final models explained 52% and 27% of the variance compared with unconditional Mobility and Self-Care models. However, some EMR data elements had apparent coding errors or missing data, and desired elements from acute care were not available. Also, unbalanced outcome data may have biased trajectories.

We demonstrate the feasibility of developing individual-level prognostic models from EMR data; however, some data elements were poorly defined, subject to error, or missing for some or all cases. Development of prognostic models from EMR will require improvements in EMR data collection and standardization.