Abstract Body: Background: Risk factor control and medication adherence are critical for stroke secondary prevention, but remain a significant challenge after discharge. We’ve developed an artificial intelligence (AI)-based algorithm to predict poor compliance to prescribed medication (PoorC-med) 90 days post-hospitalization.
Methods: Consecutive stroke patients discharged from 5 comprehensive stroke centers followed by a multimodal holistic follow-up, including a mobile app for patient communication were evaluated. PoorC-med was defined by a score >0 on the Morisky Green scale. In-hospital and early follow-up multimodal variables were evaluated; those associated with PoorC-med (p<0.05 in the univariate analysis) were used to develop 2 logistic regression models, with variables available at 7 and 30 days after discharge. The models were optimized by grid search to maximize the F2 score, with 5-fold cross-validation to predict PoorC-med at 90 days. A subsequent pool of patients following the same protocol was used for external validation.
Results: From January 1, 2020, 3261 patients were included in the multimodal follow-up; data on treatment compliance and >90 days follow-up were available for 1946 (59.7%). Of these, patients enrolled through September 23 (1801) were used to develop the AI algorithm; from October 2023, 145 patients were included in the validation set. Three hundred fifteen (17.5%) patients in the training and 33 (22.8%) in the validation set showed PoorC-med at 90 days. Variables associated with PoorC-med are shown in Fig.1. The logistic regression models (Fig. 2) showed the following performance on the training set: Confusion Matrix: [[549 937], [27 288]], Accuracy: 0.46, AUC: 0.64, F1 Score: 0.37, Recall: 0.91, Precision: 0.24, AUC PR: 0.36, AUROC: 0.72.The validation with an independent dataset yielded: Confusion Matrix: [[52 60], [3 30]], Accuracy: 0.57, AUC: 0.69, F1 Score: 0.49, Recall: 0.91, Precision: 0.33, AUC PR: 0.52, AUROC: 0.80 (Fig. 3).
Predictions using variables available only 7 days after discharge showed: Accuracy 0.45, AUC 0.63, Recall 0.92, Precision 0.23, AUROC 0.66
Conclusion: Our models are able to moderately predict poor medication compliance in stroke patients 90 days after discharge. Early identification of poorC-med patients may facilitate targeted interventions and improve secondary prevention. Further research is warranted to improve our performance and to translate the implementation of predictive models into clinical practice.