Transparently Predicting Therapy Compliance of Young Adults Following Ischemic Stroke

For stroke rehabilitation, therapy adherence is crucial to maximise patients’ recovery and stimulate re-engagement with normal daily activities. Early predictions generated by ML models that could identify ‘at risk’ patients with low levels of therapy compliance can help clinicians in adjusting therapy plans to fit patients’ capacities or procuring additional support to help patients fulfil their therapy plans. To enhance trust and transparency among patients and clinicians, these predictions need to provide enough information explaining the inner rationale of the model that helps interpret the outcomes. This paper presents a ML model trained to predict a patient’s level of compliance with a computer-based cognitive therapy rehabilitation plan. The influence of all predictive factors is analysed as a means of model explanation to provide a contextualised prediction. The model was trained and validated using electronic health records from 699 ischemic stroke patients (68.81% male, mean age 51.08) admitted at a clinic centre for cognitive rehabilitation. Predictive factors included demographic information and cognitive scores from 7 different standardised neuropsychology assessments for different cognitive domains (e.g., attention, memory, executive functions, and visual-spatial). For this study, the proportion of non-executed tasks (from the computer-based therapy plan) was adapted to fit a binary classification problem. The predictive factors contributions were computed using SHAP importance reports. The XGB algorithm was trained and evaluated using a stratified k-fold (k = 5) sampling approach. Best performance reported an ROC-AUC of 0.874 and a F1 score of 0.812. In addition, the Precision-Recall curve reported an Average Precision (AP) of 0.67. The feature importance analysis identified the age of patients and the NIHSS score as important predictive factors; however, the analysis also identified standardised cognitive assessments like the Digit Span Backwards (memory) and Images WAIS-III (visual) as relevant features to predict therapy compliance. Results show that ML-based models can assist clinicians in identifying patients with poor adherence to rehabilitation therapy. Feature importance reports revealed some of the key factors that influenced such predictions. These findings could aid clinicians in further exploring these cognitive assessments and targeting specific measured features. Additional validation is required over a larger cohort of patients to investigate the model’s accuracy and the predictive influence of some of these features.