Computerised patient-specific prediction of the recovery profile of upper limb capacity within stroke services: the next step

Abstract

Introduction: Predicting upper limb capacity recovery is important to set treatment goals, select therapies and plan discharge. We introduce a prediction model of the patient-specific profile of upper limb capacity recovery up to 6 months poststroke by incorporating all serially assessed clinical information from patients.

Methods: Model input was recovery profile of 450 patients with a first-ever ischaemic hemispheric stroke measured using the Action Research Arm Test (ARAT). Subjects received at least three assessment sessions, starting within the first week until 6 months poststroke. We developed mixed-effects models that are able to deal with one or multiple measurements per subject, measured at non-fixed time points. The prediction accuracy of the different models was established by a fivefold cross-validation procedure.

Results: A model with only ARAT time course, finger extension and shoulder abduction performed as good as models with more covariates. For the final model, cross-validation prediction errors at 6 months poststroke decreased as the number of measurements per subject increased, from a median error of 8.4 points on the ARAT (Q1-Q3:1.7-28.1) when one measurement early poststroke was used, to 2.3 (Q1-Q3:1-7.2) for seven measurements. An online version of the recovery model was developed that can be linked to data acquisition environments.

Conclusion: Our innovative dynamic model can predict real-time, patient-specific upper limb capacity recovery profiles up to 6 months poststroke. The model can use all available serially assessed data in a flexible way, creating a prediction at any desired moment poststroke, stand-alone or linked with an electronic health record system.

Keywords: models, biostatistics, biomarkers; outcome measure; prognosis; stroke; stroke unit, upper extremity.

Figure 1

ARAT recovery profiles of all 450 patients (in grey) and four typical examples in bold and the dots indicating the exact measurement points. It can be seen that the recovery of upper extremity capacity is extremely diverse, both in terms of onset ARAT score and the change over time. Most measurements were taken approximately in the first 30 days; as a result, the change after this time point can be modelled less precisely. ARAT, Action Research Arm Test.

Figure 2

Cross-validation time-dependent accuracy of the shoulder abduction and finger extension (SAFE) model for predicting 6 months ARAT score. The accuracy was defined as the absolute difference between the predicted ARAT score at 6 months poststroke from the cross-validation and the measured ARAT score at the same time and displayed as median, (IQR: Q1=25th percentile and Q3=75th percentile), lower whisker presented as Q1–1.5 * IQR and upper whisker presented as Q3 +1.5 * IQR. The accuracy is displayed as a function of the number of serial measurements were used for predicting the outcome at 6 months, of which generally the last measurement was performed at 6 months poststroke, one at 3 months poststroke and the others between 2 days and 6 weeks poststroke. ARAT, Action Research Arm Test.

Figure 3

Cross-vlidation accuracy of the fivefold cross-validation in for different levels of baseline (early poststroke) ARAT scores (left, middle and right panel) for the safe model as a function of the time of the last observed outcome (instead of the number of measurements used). Furthermore, we categorised the results by baseline ARAT group. The accuracy was defined as the difference between the predicted ARAT score at 6 months poststroke from the cross-validation and the measured ARAT score at the same time and displayed as median, (IQR: Q1=25th percentile and Q3=75th percentile), lower whisker presented as Q1–1.5 * IQR and upper whisker presented as Q3 +1.5 * IQR. The results can be different from figure 3 since some patients might have one measurement during the first week while other patients might have 2–3 repeated measurements in the first 6 months. ARAT, Action Research Arm Test.

Figure 4

Typical examples of the predicted ARAT recovery for two patients. The dotted vertical line represents the time of the last follow-up. The circles represent all the ARAT measurements available for that patient until that specific moment, while the solid line represents the predicted ARAT recovery. The shaded areas indicate the 68% (lighter shade) and 95% (darker shade) prediction intervals. from a clinical perspective, the errors in the cross-validation provide the best estimate of what the error in predicting the outcome for a new patient will be and may, therefore, be most clinically relevant. For each patient, the predicted recovery is illustrated at a first and a second time point, not necessarily corresponding to the first and second available measurements from a patient. The data of the same patients can be downloaded in the online APP to visualised predictions at all time points: https://emcbiostatistics.shinyapps.io/DynamicPredictionARATapp https://emcbiostatistics.shinyapps.io/DynamicPredictionARATapp/. ARAT, Action Research Arm Test.