Background and Aims: Artificial Intelligence-Machine Learning (AI-ML) models were developed using the data from Indian Registry in Ischemic Stroke tenecteplase (IRIS-TNK). The aim of this study is to perform a comparative assessment of most used AI-ML models to predict the long-term clinical outcome of stroke patients treated with tenecteplase, using patients’ baseline characteristics.
Methods: In IRIS-TNK registry, 1015 patients with median age of 62 (inter-quartile range of 52-71) years, were recruited across 20 cites in India. Data was checked for imbalance and class weights and random over sampling were used to deal with the imbalance in data. No imputation was done. The clinical outcome was measured in terms improvement in modified Rankin Scale (mRS) score from day 7 to day 90, post tenecteplase administration.
Ten most popular multiclass classification models were trained using the registry data and base models were built. Out of these, Random Forest (RF), Gradient Boosting (GB), XG Boost, and CatBoost algorithms gave promising results with respect to model performance metrics. The performance of these four models were further improved by hyperparameter tuning.
Results: RF and GB models gave best results than the rest wherein RF gave an accuracy of 82.22%, precision of 80.97%, recall of 82.22% and a F1-score of 75.16% and GB gave accuracy of 80.00%, precision of 73.12%, recall of 80.00% and a F1-score of 73.07%.
Conclusion: The results demonstrated that AL-ML model can achieve significant predictive accuracy, indicating its potential utility in clinical settings to aid in the prognosis.