Accuracy of an Algorithm in Predicting Upper Limb Functional Capacity in a United States Population

Objective: To determine the accuracy of an algorithm, using clinical measures only, on a sample of persons with first-ever stroke in the United States (US). It was hypothesized that algorithm accuracy would fall in a range of 70%-80%.

Design: Secondary analysis of prospective, observational, longitudinal cohort; 2 assessments were done: (1) within 48 hours to 1 week poststroke and (2) at 12 weeks poststroke.

Setting: Recruited from a large acute care hospital and followed over the first 6 months after stroke.

Participants: Adults with first-ever stroke (N=49) with paresis of the upper limb (UL) at ≤48 hours who could follow 2-step commands and were expected to return to independent living at 6 months.

Intervention: Not applicable.

Main outcome measures: The overall accuracy of the algorithm with clinical measures was quantified by comparing predicted (expected) and actual (observed) categories using a correct classification rate.

Results: The overall accuracy (61%) and weighted κ (62%) were significant. Sensitivity was high for the Excellent (95%) and Poor (81%) algorithm categories. Specificity was high for the Good (82%), Limited (98%), and Poor (95%) categories. Positive predictive value (PPV) was high for Poor (82%) and negative predictive value (NPV) was high for all categories. No differences in participant characteristics were found between those with accurate or inaccurate predictions.

Conclusions: The results of the present study found that use of an algorithm with clinical measures only is better than chance alone (chance=25% for each of the 4 categories) at predicting a category of UL capacity at 3 months post troke. The moderate to high values of sensitivity, specificity, PPV, and NPV demonstrates some clinical utility of the algorithm within health care settings in the US.