Clinicians often base their predictions of walking and arm recovery on multiple predictors. Multivariate prediction models may assist clinicians to make accurate predictions. Several reviews have been published on the prediction of motor recovery after stroke, but none have critically appraised development and validation studies of models for predicting walking and arm recovery.
In this review, we highlight some common methodological limitations of models that have been developed and validated. Notable models include the proportional recovery model and the PREP algorithm. We also identify five other models based on clinical predictors that might be ready for further validation. It has been suggested that neurophysiological and neuroimaging data may be used to predict arm recovery. Current evidence suggests, but does not show conclusively, that the addition of neurophysiological and neuroimaging data to models containing clinical predictors yields clinically important increases in predictive accuracy.
It would be useful to be able to predict recovery of walking and arm after stroke. Accurate predictions are needed so that clinicians can provide patients with prognoses, set goals, select therapies and plan discharge [1,2,3,4]. For example, if it was possible to predict with some certainty that a particular patient would be unable to walk independently at six months, the clinicians providing that patient with acute and subacute care might work toward a discharge goal of safe transfers. Therapy might involve carer training and equipment prescription rather than intensive gait training. The ability to make accurate predictions could reduce the length of stay in hospitals and enable efficient utilization of stroke care resources [4,5].
Several systematic reviews have identified strong predictors of walking and arm recovery after stroke [2,3,6]. In one systematic review of prognostic studies on walking, clinical variables such as age, severity of paresis and leg power were found to be strong predictors of walking after stroke (based on five studies, each of between 197 and 804 patients) . In another systematic review of prognostic studies on arm recovery, clinical, neurophysiological and neuroimaging data were found to be strong predictors of arm recovery after stroke (based on 58 studies of 9–1197 patients) . These clinical, neurophysiological and neuroimaging data included measures of upper limb impairment, upper limb function, lower limb impairment, motor and somatosensory evoked potentials, and measures obtained with diffusion tensor imaging .
In practice, clinicians base their predictions about clinical outcomes on multiple variables [7,8,9]. If multiple predictors are to be used to make prognoses, there needs to be a proper accounting of the independent (incremental) predictive value of each predictor variable. Therefore the most useful information about prognosis is likely to come from multivariate prediction models [7,8,9].
The research which underpins establishment of clinically useful multivariate prediction models involves several steps. First ‘development studies’ are conducted to build the multivariate prediction models . Subsequently the predictive accuracy of the models is tested on new cohorts [7,10]. These studies are known as ‘validation studies’ . It is recommended that prediction models should not be used in clinical practice until both development and validation studies have been conducted [7,10]. Once development and validation studies have been conducted, impact studies may be conducted, although the reality is that few reports of impact studies are published. Impact studies resemble clinical trials; they test the efficacy of use of prediction models on patient outcomes, clinician behaviour and cost-effectiveness of care [7,11]. Recent narrative reviews have provided updates on the prediction of motor recovery after stroke [5,12] but these reviews have not focused on development and validation studies of models for predicting walking and arm recovery.
This review provides a critical review of prediction models of walking and arm recovery after stroke. Studies were identified using the search strategy and inclusion criteria in the Appendix. The review begins in the second section with the definitions and measurements of walking and arm recovery. The third section provides a detailed description of the recommended process for developing and validating a prediction model because this process provides a benchmark against which prediction modelling studies of walking and arm recovery can be evaluated. The fourth section critically appraises development and validation studies of walking and arm recovery with the aim of identifying multivariate models that could potentially be implemented in clinical practice. Much has been written about the role of neurophysiological and neuroimaging data in predicting arm recovery. The fifth section considers whether neurophysiological and neuroimaging data provide additional predictive value over clinical data alone in predicting arm recovery. We conclude with a summary and recommendations for future prediction modelling studies.